Intelligent early warning method and system for childbirth monitoring in response to abnormal heart rate of fetus

Abstract

The invention relates to a delivery monitoring intelligent early warning method and system responding to fetal heart rate abnormity, and relates to the technical field of delivery monitoring early warning, and the method comprises the steps: obtaining delivery monitoring information, medical record information and real-time drug administration information of a target puerpera; forward predicting the distress state of the fetus through a first prediction model, and obtaining a real-time distress state vector; constructing a second prediction model, reversely predicting the distress state of the fetus, and obtaining subjective distress state vector distribution; and according to the distribution of the real-time distress state vector and the subjective distress state vector, the monitoring abnormality is calculated, and delivery monitoring early warning is carried out according to an early warning threshold. The problems of one-sided risk assessment and insufficient early warning accuracy and timeliness caused by incomplete information integration and excessive dependence on manual judgment of traditional childbirth monitoring early warning are solved.

Description

Technical Field

[0001] This application relates to the field of labor monitoring and early warning, and in particular to intelligent early warning methods and systems for labor monitoring in response to fetal heart rate abnormalities. Background Technology

[0002] With the increasing demands for maternal and infant safety in obstetric medicine, the timely identification and early warning of fetal heart rate abnormalities during delivery has become a key technical requirement to assist medical staff in accurately assessing the fetal condition.

[0003] Currently, traditional labor monitoring and early warning methods do not fully integrate various key information affecting fetal status and rely excessively on human experience and judgment. This not only makes risk assessments prone to bias due to missing information or subjective bias, but also reduces the accuracy and timeliness of early warnings, increases the uncertainty of intervention decisions in the delivery room, and makes it difficult to meet the clinical needs for precise and efficient labor monitoring. Summary of the Invention

[0004] This application provides a method and system for intelligent early warning of labor monitoring in response to fetal heart rate abnormalities, which improves the current situation of traditional labor monitoring and early warning, which suffers from incomplete information integration and excessive reliance on manual judgment, resulting in one-sided risk assessment and insufficient accuracy and timeliness of early warning.

[0005] The embodiments of this application disclose the following technical solutions: In a first aspect, embodiments of this application provide a method for intelligent early warning of labor monitoring in response to fetal heart rate abnormalities, the method comprising: Acquire the labor monitoring information, medical record information, and real-time drug administration information of the target pregnant woman, wherein the labor monitoring information includes at least fetal heart rate timing data and uterine contraction timing data; Based on the labor monitoring information and the medical record information, the fetal distress state is positively predicted through the first prediction model to obtain a real-time distress state vector; A second prediction model is constructed, and the medical record information and the real-time drug administration information are input into the second prediction model to predict the fetal distress state in reverse and obtain the subjective distress state vector distribution. Based on the distribution of the real-time distress state vector and the subjective distress state vector, the monitoring abnormality degree is calculated, and a labor monitoring early warning based on the preset early warning threshold is performed.

[0006] Secondly, embodiments of this application provide an intelligent early warning system for labor monitoring in response to fetal heart rate abnormalities, the system comprising: The maternal multi-source information acquisition module is used to acquire the target maternal labor monitoring information, medical record information, and real-time drug administration information, wherein the labor monitoring information includes at least fetal heart rate timing data and uterine contraction timing data; The real-time distress positive prediction module is used to predict the fetal distress state positively through the first prediction model based on the labor monitoring information and the medical record information, and obtain the real-time distress state vector. The second model reverse prediction module is used to construct a second prediction model and input the medical record information and the real-time drug administration information into the second prediction model to reverse predict the fetal distress state and obtain the subjective distress state vector distribution. The monitoring anomaly calculation and early warning module is used to calculate the monitoring anomaly degree based on the distribution of the real-time distress state vector and the subjective distress state vector, and to provide early warning for labor monitoring based on the preset early warning threshold.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application proposes an intelligent early warning method and system for fetal heart rate monitoring in response to fetal heart rate abnormalities. By collecting multi-source information of the mother in stages, positively predicting the real-time distress state, constructing and applying a second prediction model to reversely deduce the risk distribution, calculating the degree of monitoring abnormality and triggering precise early warning, intelligent early warning of fetal heart rate abnormalities during labor is realized. First, labor monitoring information, medical records, and real-time medication information of the target mothers are collected to provide comprehensive data support for subsequent prediction and early warning. Then, based on the labor monitoring and medical records, a first prediction model is used to positively predict fetal distress status, outputting a real-time distress status vector reflecting the current fetal state. Next, a second prediction model is constructed, and the medical records and real-time medication information are input into the second prediction model to perform a preset number of iterative predictions. A multi-dimensional vector space is constructed and mapped to the distress status vector to obtain the subjective distress status vector distribution. The empirical confidence level and subjective confidence level are calculated based on the relative density distribution and voting confidence assessment results. Finally, the integrated prediction output is selected as the reference distress status vector group. Historical fetal normal status vectors are obtained, and the global covariance matrix is ​​calculated. The Mahalanobis distance between the real-time distress status vector and each reference vector is calculated using the matrix. Multiple Mahalanobis distances are weighted with empirical confidence and subjective confidence as weights to obtain the monitoring abnormality level. This is compared with a preset early warning threshold to trigger a labor monitoring early warning based on medication information.

[0008] The technical solution proposed in this application solves the problems in traditional labor monitoring, such as low efficiency due to reliance on manual interpretation, one-sided risk assessment due to reliance on a single real-time data, and lack of accurate numerical basis and clear intervention direction for early warning. It avoids the delay and misjudgment of early warning caused by differences in human experience or incomplete information coverage, and improves the accuracy and timeliness of early warning of abnormal fetal heart rate. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart illustrating the intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities provided in this application embodiment; Figure 2 This is a schematic diagram of the intelligent early warning system for labor monitoring in response to fetal heart rate abnormalities provided in an embodiment of this application.

[0011] The components represented by each number in the attached diagram are explained below: The system includes a multi-source information acquisition module for parturients (01), a real-time distress positive prediction module (02), a second model reverse prediction module (03), and a monitoring abnormality calculation and early warning module (04). Detailed Implementation

[0012] This application provides a method and system for intelligent early warning of labor monitoring in response to fetal heart rate abnormalities, which solves the technical problems of traditional labor monitoring and early warning in the prior art, which are due to incomplete information integration and excessive reliance on manual judgment, resulting in one-sided fetal risk assessment and insufficient accuracy and timeliness of early warning.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0015] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0016] Example 1, as shown in the appendix Figure 1 As shown, this application provides an intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities, the method comprising the following steps: S110: Obtain the labor monitoring information, medical record information, and real-time drug administration information of the target pregnant woman, wherein the labor monitoring information includes at least fetal heart rate timing data and uterine contraction timing data; In this embodiment of the application, in the scenario of real-time monitoring and intelligent early warning of fetal heart rate abnormalities during labor, in order to fully grasp the key information affecting the fetal state, it is necessary to first obtain the target mother's labor monitoring information, medical record information, and real-time drug administration information to ensure that the data can cover the fetal real-time physiological signals, the mother's individual health background, and the current medical intervention, so as to provide reliable input data for the positive prediction and reverse deduction of the subsequent fetal distress state.

[0017] Firstly, regarding the acquisition of labor monitoring information, the focus should be on collecting fetal heart rate timing data and uterine contraction timing data, which directly reflect the real-time physiological state of the fetus. Fetal heart rate timing data is continuously collected using specialized fetal heart rate monitoring equipment, recording the dynamic changes in fetal heart rate in a time-series format.

[0018] For example, starting from the beginning of the active labor phase, the fetal heart rate value is continuously recorded every unit of time, and the fluctuation of the fetal heart rate before and after the occurrence of uterine contractions and during the duration of uterine contractions is captured simultaneously. When the mother has regular uterine contractions, it is necessary to clearly record whether the fetal heart rate is accompanied by deceleration, acceleration or abnormal variation. These data are the key basis for judging whether the fetus has intrauterine hypoxia and tolerance to labor stress.

[0019] In addition, contraction timing data is obtained through a contraction monitoring device, which records the occurrence time, duration, interval, and intensity changes of contractions. For example, it records the start and end times of a contraction, the interval between two contractions, and the peak pressure during a contraction. At the same time, the contraction data is correlated with the fetal heart rate data of the corresponding time period, providing temporal data support for subsequent analysis of the correlation between contractions and fetal heart rate changes.

[0020] Furthermore, when obtaining the medical records of the target pregnant woman, it is necessary to retrieve key information that reflects the pregnant woman's pregnancy status and individual health condition, mainly by extracting the pregnant woman's prenatal check-up records and past health records from the hospital's electronic medical record platform.

[0021] Specifically, medical record information mainly includes the mother's gestational age, age, previous pregnancy history, pregnancy complications, and fetal position. For example, if the mother has complications such as gestational hypertension or gestational diabetes, or if the fetus has abnormal fetal position or low amniotic fluid, this information needs to be obtained completely, because such factors increase the risk of fetal heart rate abnormalities during delivery. Subsequent predictive models need to adjust their judgments based on medical record data to avoid inaccurate risk assessments due to ignoring individual differences.

[0022] Finally, the acquisition of real-time drug administration information needs to be synchronized with the labor process, recording all details of drug interventions received by the mother during labor, which should be entered and updated in real time by medical staff through medical recording devices.

[0023] Specifically, real-time drug administration information includes the name of the drug administered, the time of administration, the dosage, the route of administration, and subsequent dosage adjustments. For example, when a pregnant woman needs to use medication to strengthen uterine contractions due to uterine atony, the initial dosage, the time of administration, the timing of dosage adjustments based on the uterine contraction situation, and the adjusted dosage should be recorded.

[0024] The three types of information obtained through the above methods provide complete data support from three aspects: real-time fetal physiological signals, maternal health history, and current medical interventions. This avoids the one-sidedness of information caused by traditional monitoring relying solely on single fetal heart rate data. It provides comprehensive input data for the subsequent positive prediction of the real-time distress state vector by the first prediction model and the reverse derivation of the subjective distress state vector distribution by the second prediction model.

[0025] S120: Based on the labor monitoring information and the medical record information, the fetal distress state is positively predicted through the first prediction model to obtain a real-time distress state vector; In this embodiment of the application, in order to quickly determine whether the fetus is at risk of distress based on the acquired real-time physiological signals of the fetus and the individual health background of the mother, the delivery monitoring information and medical record information need to be input into the first prediction model for positive prediction, so as to output a real-time distress state vector that quantifies the fetal distress state.

[0026] Specifically, the first prediction model, as an existing model applicable to the classification and prediction of multi-source medical data, adopts a machine learning model based on feature engineering. The model has been trained and validated using a large amount of historical delivery data (including different monitoring information, medical record information and corresponding fetal distress diagnosis results). It has the ability to map input data into a fetal distress state vector and can adapt to the processing efficiency requirements of real-time data in the delivery scenario.

[0027] During the model input phase, the delivery monitoring information and medical record information need to be processed for data adaptation to ensure that the two types of data meet the input format requirements of the first prediction model.

[0028] Specifically, for fetal heart rate time series data in labor monitoring information, it is necessary to extract its key feature values, such as the fetal heart rate baseline per unit time, fetal heart rate variability, whether deceleration occurs and the type of deceleration, etc. For uterine contraction time series data, features such as uterine contraction frequency, average uterine contraction duration, and peak uterine contraction intensity are extracted.

[0029] For medical record information, non-numerical data needs to be converted into numerical codes, and numerical data needs to be normalized, such as mapping gestational age data to the 0-1 range, to ensure that the feature magnitude is consistent with that of monitoring information and to avoid affecting the model prediction accuracy due to differences in data scale.

[0030] Furthermore, the processed labor monitoring information features and medical record information features are fused into a unified input feature vector, which is then input into the first prediction model for positive prediction.

[0031] The aforementioned positive prediction model, based on real-time and historical input data, directly outputs the current fetal distress status. In other words, the model analyzes and judges the current input data based on the correlation between features learned during training and distress status.

[0032] For example, if the input fetal heart rate time series data shows that the baseline fetal heart rate is consistently below 110 beats / minute, accompanied by late decelerations, and the medical record information shows that the mother has a history of gestational diabetes and the fetus is in breech presentation, the model will combine the combined effects of these characteristics to determine that the fetus is currently at moderate risk of distress. If the baseline fetal heart rate is stable at 130-140 beats / minute, the fetal heart rate variability is normal, there are no decelerations, and the medical record information shows that the mother has no pregnancy complications and the fetal position is normal, the model will determine that the fetus is currently in a normal state.

[0033] Ultimately, the results output by the first prediction model are presented in the form of a real-time distress state vector. The dimensions of this vector correspond to the assessment dimensions of the fetal distress state. For example, the real-time distress state vector may include three dimensions: fetal heart rate score, uterine contraction tolerance score, and basic health risk score. The value of each dimension corresponds to the degree of fetal risk under different dimensions. The vector as a whole completely quantifies the current fetal distress state.

[0034] For example, a real-time distress state vector of [3, 2, 1] represents a fetal heart rate score of 3 (low risk), a uterine contraction tolerance score of 2 (low risk), and a basic health risk score of 1 (low risk), reflecting that the fetus is currently not at risk of distress. If the vector is [7, 6, 5], it indicates that all dimensions are in the medium to high risk range, suggesting that the fetus may be in distress.

[0035] Through the positive prediction of the first prediction model, the originally scattered and complex labor monitoring information and medical record information can be transformed into an intuitive and quantifiable real-time distress state vector. This lays a real-time data foundation for combining the subjective distress state vector distribution obtained from the subsequent reverse prediction to carry out further anomaly calculation and early warning judgment.

[0036] S130: Construct a second prediction model and input the medical record information and the real-time drug administration information into the second prediction model to predict the fetal distress state in reverse and obtain the subjective distress state vector distribution; In this embodiment of the application, in order to avoid early warning deviations caused by insufficient risk prediction, it is necessary to first construct a second prediction model with reverse derivation capability, and then input medical record information and real-time drug administration information into the model to carry out reverse prediction in order to obtain a subjective distress state vector distribution that can reflect multiple potential situations of fetal distress.

[0037] First, the construction of the second prediction model needs to be completed. Specifically, the inverse analysis indicator set is first determined, and then samples are collected based on the inverse analysis indicator set to obtain inverse analysis sample data containing sample inverse indicator vectors and corresponding sample distress state vectors, providing basic data that meets the requirements of inverse prediction for model training.

[0038] Furthermore, using the reverse analysis sample data as training data, multiple machine learning-based weak mappers are constructed and trained. Each weak mapper uses the sample reverse indicator vector as the mapping input and the sample distress state vector as the mapping supervision, allowing each weak mapper to independently learn the correlation between the reverse indicator and the distress state.

[0039] Furthermore, multiple weak mappers are integrated by voting based on the ensemble learning method to form an ensemble prediction core that can integrate the judgment results of multiple weak mappers. The ensemble prediction core is then connected to the voting confidence evaluation layer. The voting confidence evaluation layer is used to conduct a preliminary assessment of the credibility of the prediction results. Finally, the construction of the second prediction model is completed, enabling the model to reverse-engineer the fetal distress state and assess the credibility of the prediction results.

[0040] After completing the construction of the second prediction model, the medical record information and real-time drug administration information are input into the model, and a preset number of iterative predictions are performed. During the iteration process, the integrated prediction output generated by each prediction and the corresponding voting confidence assessment results are obtained simultaneously. Through multiple iterations, more potential fetal distress conditions are covered.

[0041] Furthermore, a multidimensional vector space is constructed, and the distress state vectors in all the integrated prediction outputs obtained through iteration are mapped one by one into this multidimensional vector space to form a subjective distress state vector distribution that can intuitively present the distribution characteristics of various potential distress states.

[0042] Furthermore, the relative density distribution of the subjective distress state vector distribution is calculated, and an empirical confidence level is defined for each point in the vector distribution based on the relative density distribution, so as to reflect the data support strength of each point in the distribution. At the same time, combined with the previously obtained voting confidence assessment results, the subjective confidence level of each point in the subjective distress state vector distribution is calculated accordingly, so as to reflect the model output credibility of the prediction results of each point.

[0043] Finally, the distribution of the subjective distress state vector is correlated with the calculated subjective confidence level and output to ensure that subsequent analysis can not only grasp the various potential situations of fetal distress, but also clarify the credibility of each situation, providing more comprehensive and reliable input data for subsequent monitoring abnormality calculation.

[0044] Step S140 in the method provided in this application embodiment includes: Determine the reverse analysis indicators, and collect samples based on the reverse analysis indicators to obtain reverse analysis sample data, wherein the reverse analysis sample data includes at least the sample reverse indicator vector and the corresponding sample distress state vector. Using the reverse analysis sample data as training data, multiple machine learning-based weak mappers are constructed and trained, wherein the weak mappers use the sample reverse indicator vector as the mapping input and the sample distress state vector as the mapping supervision. The multiple weak mappers are integrated by voting using an ensemble learning method to obtain an ensemble prediction core, and the ensemble prediction core is connected to a voting confidence evaluation layer to obtain the second prediction model.

[0045] Input the medical record information and the real-time drug administration information into the second prediction model, perform a preset number of iterative predictions, and obtain multiple integrated prediction outputs and corresponding voting confidence assessment results; Construct a multidimensional vector space and map the distress state vectors from multiple integrated prediction outputs to the multidimensional vector space to obtain the distribution of the subjective distress state vectors. Calculate the relative density distribution of the subjective distress state vector distribution, and define the empirical confidence level of each point in the subjective distress state vector distribution based on the relative density distribution; Based on the voting confidence assessment results, the subjective confidence level of each point in the subjective distress state vector distribution is calculated accordingly; The subjective distress state vector distribution is correlated with the subjective confidence level.

[0046] In this embodiment of the application, in order to comprehensively analyze the potential impact of medical record information and real-time drug administration information on fetal distress through reverse derivation, it is necessary to construct a second prediction model through multiple stages and carry out reverse prediction to obtain a subjective distress state vector distribution that combines multiple possibilities and credibility assessment, so as to provide a basis for subsequent reliable labor monitoring and early warning.

[0047] Specifically, the first step is to determine the inverse analysis indicators and collect inverse analysis sample data to ensure that the subsequent second predictive model can be trained based on accurate inputs that closely align with clinical practice. The inverse analysis sample data must include at least the sample inverse indicator vector and the corresponding sample distress state vector.

[0048] The method provided in this application embodiment includes determining a reverse analysis indicator set and collecting samples based on the reverse analysis indicator set to obtain reverse analysis sample data, including: Based on prior knowledge of fetal distress, relevant indicators are obtained, and through association analysis, a set of association coefficients between the relevant indicators and fetal distress is obtained. Based on the medical record information and the real-time drug administration information, extract the current status indicators and take the intersection of the current status indicators and the relevant indicators; Normalize the set of correlation coefficients and calculate the cumulative correlation coefficients corresponding to the intersection; If the cumulative correlation coefficient is less than the preset cumulative limit, the intersection is output as the reverse analysis index; otherwise, the intersection is serialized and the first N intersection factors that satisfy the cumulative limit are selected and output as the reverse analysis index.

[0049] Specifically, the first step is to obtain relevant indicators based on prior knowledge of fetal distress. This prior knowledge of fetal distress includes various indicators that have been validated in obstetric clinical practice and are associated with fetal distress, such as maternal pregnancy complications, fetal position, type and dosage of medication, and baseline variability of fetal heart rate. These clinically validated indicators are integrated to form a preliminary set of relevant indicators, ensuring that the indicators have basic medical rationale.

[0050] Furthermore, the association coefficient set of relevant indicators to fetal distress is calculated by association analysis. The association analysis method can quantify the association strength between each indicator and the occurrence of fetal distress. For example, the higher the frequency of a certain indicator in fetal distress cases and the stronger its correlation with the time of distress, the larger the corresponding association coefficient.

[0051] Furthermore, current status indicators are extracted based on medical record information and real-time medication information. The medical record information contains the current mother's specific health background, while the real-time medication information covers the details of the medication interventions the mother received during labor. Indicators potentially related to fetal distress are extracted from this information to form a current status indicator set, ensuring that the indicators accurately match the current mother's individual situation.

[0052] For example, if the current mother has gestational diabetes and is using oxytocin, the current status index collection should include targeted indicators such as the history of gestational diabetes and the current dose of oxytocin.

[0053] Furthermore, by taking the intersection of the current condition indicators and the relevant condition indicators, we can screen out indicators that are both confirmed to be related to distress in clinical prior knowledge and exist in the current actual situation of the pregnant woman. We can then eliminate indicators that only meet prior knowledge but are irrelevant to the current pregnant woman, thus initially achieving scenario-based screening of indicators.

[0054] After obtaining the intersection, the correlation coefficient set needs to be normalized. Specifically, since the original values ​​of the correlation coefficients of different indicators may have different magnitudes, direct calculation will lead to high-magnitude correlation coefficients overly dominating the results. Normalization can map all correlation coefficients to a unified interval to ensure that the correlation strength of each indicator has an equal comparison benchmark in subsequent calculations.

[0055] Furthermore, the cumulative correlation coefficient corresponding to the intersection is calculated, that is, the normalized correlation coefficients of all indicators in the intersection are summed to determine whether the overall correlation strength of the intersection indicators is reasonable.

[0056] If the cumulative correlation coefficient is less than the preset cumulative limit, it means that the overall correlation strength of the indicators in the intersection is moderate and there is no redundancy caused by too many highly correlated indicators. In this case, the intersection is directly output as the indicator set for reverse analysis.

[0057] Conversely, if the cumulative correlation coefficient exceeds the preset cumulative limit, it indicates that there are too many highly correlated indicators in the intersection, which may be due to information overlap or over-focus on a certain type of factor. The intersection needs to be serialized from large to small according to the normalized correlation coefficient, and the top N indicators (the value of N is based on the cumulative correlation coefficient just meeting the preset cumulative limit) are selected. The output is the reverse analysis indicator set.

[0058] For example, if the preset cumulative limit is 1.5, the cumulative correlation coefficient of the first 3 indicators after serialization is 1.4, and that of the first 4 is 1.6, then the first 3 indicators are selected to ensure that the final reverse analysis indicator set covers both highly correlated core indicators and avoids redundant interference.

[0059] After determining the inverse analysis indices, samples are collected based on these indices to obtain inverse analysis sample data. Each sample must contain a sample inverse indices vector corresponding to the inverse analysis indices, as well as a sample distress state vector for each vector. This ensures that the sample data perfectly matches the dimensions of the inverse analysis indices, providing suitable supervision data for subsequently building and training multiple machine learning-based weak mappers.

[0060] Furthermore, using the reverse analysis sample data as training data, multiple machine learning-based weak mappers were constructed and trained. Among these, decision trees were chosen as the machine learning model for the weak mappers. The decision tree model associates sample features with prediction results through a tree-like branching structure, intuitively presenting the correspondence between reverse analysis indicators and fetal distress states. It also boasts high training efficiency and strong adaptability to real-time delivery monitoring scenarios.

[0061] During training, each weak mapper takes the sample inverse indicator vector as the mapping input, such as the vector input model formed by combining "history of gestational diabetes (coded value) + current dose of oxytocin (normalized value) + fetal position (coded value)".

[0062] Meanwhile, the sample distress state vector is used as the mapping supervision. This vector can be obtained by quantifying the clinical diagnosis results, such as the numerical vectors of "no distress (0), mild distress (1), moderate distress (2), and severe distress (3)", so that the weak mapper can gradually learn the mapping rules between the reverse indicator combination and the distress state during training.

[0063] During training, cross-validation of each weak mapper is also required. For example, 5-fold cross-validation can be used to divide the training set and validation set, and the model parameters can be adjusted to ensure that each weak mapper has basic prediction ability on the validation set, and to avoid excessive prediction bias caused by improper parameters of a single weak mapper.

[0064] Furthermore, multiple weak mappers are ensembled through voting using an ensemble learning method to obtain the ensemble prediction core. Ensemble learning reduces the bias and variance of a single model and improves overall prediction stability by combining the judgment results of multiple weak mappers.

[0065] Specifically, firstly, the reverse indicator vectors of the same set of samples are input into all trained weak mappers, and the prediction results of the sample distress state vectors output by each weak mapper are collected. Then, according to the majority voting rule, the number of votes received by each distress state vector is counted, and the distress state vector with the most votes is selected as the ensemble prediction result, forming the core of the ensemble prediction.

[0066] For example, if 6 out of 10 weak mappers predict "mild distress (1)", 3 out of 3 out of 10 ...

[0067] Furthermore, the acquired integrated prediction core is connected with the voting confidence assessment layer to construct a second prediction model, which simultaneously possesses the ability to reverse predict fetal distress and the ability to assess the credibility of prediction results.

[0068] The method provided in this application includes the following execution steps for the voting confidence evaluation layer: Obtain the integrated prediction output of the integrated prediction core, wherein the integrated prediction output includes multiple distress state vectors and corresponding voting ratios; Obtain the confidence scores of multiple weak mappers corresponding to the distress state vector, calculate the mean confidence score, calculate the product of the mean confidence score and the voting ratio, and output the voting confidence score. The voting confidence scores corresponding to multiple distress state vectors are obtained through iteration, and the output is the voting confidence evaluation result.

[0069] In this embodiment of the application, in order to avoid relying solely on the voting ratio of weak mappers to determine the reliability of the fetal distress prediction results, which could lead to the overall credibility assessment being affected by the insufficient performance of some weak mappers, it is necessary to correct the integrated prediction output through a voting confidence assessment layer to quantify the credibility of the prediction results of each distress state vector, so as to further ensure the accuracy of labor monitoring and early warning.

[0070] Specifically, the first step is to obtain the ensemble prediction output of the ensemble prediction core. The ensemble prediction core is formed by ensemble voting across multiple machine learning-based weak mappers. Its output includes multiple distressed state vectors and the voting proportion for each vector. The voting proportion represents the percentage of votes received by the distressed state vector across all weak mapper predictions relative to the total number of votes.

[0071] For example, for a given input of medical record information and real-time drug administration information, the integrated prediction core outputs three distress state vectors, corresponding to no distress, mild distress, and moderate distress, respectively. The mild distress vector receives 6 votes out of 10 weak mappers, with a voting ratio of 0.6; the no distress vector receives 3 votes, with a voting ratio of 0.3; and the moderate distress vector receives 1 vote, with a voting ratio of 0.1. The aforementioned distress state vectors and their corresponding voting ratios together constitute the integrated prediction output.

[0072] Furthermore, the confidence scores of multiple weak mappers corresponding to each distressed state vector are obtained, and the mean confidence score is calculated. Each weak mapper generates a confidence score simultaneously when outputting its distressed state vector prediction result. This confidence score reflects the degree of confidence the weak mapper has in its own prediction results, and its magnitude is positively correlated with the prediction accuracy of the weak mapper on similar samples during training.

[0073] For example, if a weak mapper has a prediction accuracy of 85% for mildly distressed samples in its historical training, then the confidence level for its output of a mildly distressed vector this time can be set to 0.85.

[0074] Furthermore, for a specific distress state vector in the integrated prediction output, it is necessary to collect the confidence scores of all weak mappers that predict that vector. For example, the confidence scores of the six weak mappers predicting mild distress are 0.8, 0.85, 0.75, 0.9, 0.82, and 0.78, respectively. These confidence scores are added together and averaged, i.e., (0.8+0.85+0.75+0.9+0.82+0.78) / 6=0.82, which gives the mean confidence score of the mild distress vector.

[0075] Furthermore, the calculated mean confidence level is multiplied by the voting ratio corresponding to the mildly distressed vector to obtain the voting confidence level, thus achieving the effect of correcting the voting ratio using the performance of the weak mapper. For example, multiplying the mean confidence level of the mildly distressed vector (0.82) by the voting ratio (0.6) yields a voting confidence level of 0.492. This value reflects both the advantage of the number of votes and incorporates the predictive reliability of the weak mapper, making it more valuable than a simple voting ratio.

[0076] Finally, the voting confidence scores corresponding to multiple distressed state vectors are obtained through iteration, and the output is the voting confidence assessment result. Specifically, for each distressed state vector in the ensemble prediction output, the above steps of "obtaining confidence scores - calculating the mean - multiplying by the voting ratio" are repeated, and all distressed state vectors and their corresponding voting confidence scores are organized and integrated to form a complete voting confidence assessment result.

[0077] Among them, the voting confidence assessment results can clearly present the credibility of each predicted distress state vector, avoiding misjudgment caused by ignoring the performance differences of weak mappers when selecting vectors based solely on voting ratios. This provides credibility support for inputting medical record information and real-time drug administration information into the second prediction model to carry out reverse prediction and obtain the distribution of subjective distress state vectors.

[0078] Furthermore, during the connection process between the integrated prediction core and the voting confidence assessment layer, a data flow interaction relationship needs to be established between the integrated prediction core and the voting confidence assessment layer. Specifically, after the integrated prediction core outputs the integrated prediction output containing the distress state vector and the voting ratio, this output is automatically passed to the voting confidence assessment layer.

[0079] Furthermore, the voting confidence assessment layer calculates the voting confidence level according to a preset process, generates the voting confidence assessment result, and feeds the result back to the integrated prediction core, forming a data processing link of "prediction output - confidence assessment - result feedback". Through the above connection, the second prediction model can not only output the reverse prediction result of the fetal distress state, but also simultaneously provide the confidence index of the result, avoiding the problem of only outputting the prediction vector without being able to judge the reliability.

[0080] Furthermore, after completing the construction of the second prediction model, the patient record information and real-time drug administration information are input into the model, and a preset number of iterative predictions are performed.

[0081] The preset number of times needs to be determined in combination with the real-time requirements of labor monitoring and the requirements of comprehensive prediction. For example, it can be set to 10-15 times. During each iteration of prediction, the second prediction model will generate different integrated prediction outputs based on subtle input feature perturbations.

[0082] Simultaneously, each iteration synchronously acquires the corresponding voting confidence assessment results, ensuring that the iteration process not only covers more potential distress states but also records the confidence data for each possibility. For example, given a pregnant woman's information of "history of gestational diabetes + oxytocin 3 units / hour," in 10 iterations of prediction, 6 outputs indicate mild distress (mean voting confidence 0.48), 3 outputs indicate no distress (mean voting confidence 0.25), and 1 outputs indicate moderate distress (mean voting confidence 0.07), forming a correspondence between multiple sets of integrated prediction outputs and voting confidence assessment results.

[0083] Furthermore, a multidimensional vector space is constructed, and distress state vectors from multiple ensemble prediction outputs are mapped to this space. The dimensions of the multidimensional vector space must be consistent with the dimensions of the distress state vectors. For example, if the distress state vectors include three dimensions—fetal heart rate risk score, drug administration impact score, and basic health risk score—then the multidimensional vector space is set to 3 dimensions.

[0084] During the mapping process, the numerical values ​​of each dimension of the distress state vector are used as spatial coordinates to mark the corresponding point positions in the multidimensional vector space. For example, the mild distress vector (3, 4, 2) corresponds to the point with coordinates (3, 4, 2) in 3-dimensional space. All distress state vectors obtained by iterative prediction are mapped according to this rule, and finally form a discrete set of points in the multidimensional space, which is the distribution of subjective distress state vectors.

[0085] Among them, the obtained distribution of subjective distress state vectors can intuitively show the frequency of occurrence and spatial clustering characteristics of different distress state vectors. For example, the mapping points of mild distress vectors have a high density of clustering in space, indicating that the reverse prediction results of the second prediction model for this state are more concentrated.

[0086] Furthermore, the relative density distribution of the subjective distress state vector distribution is calculated, and the empirical confidence level of each point in the distribution is defined based on the obtained relative density distribution.

[0087] Specifically, relative density calculation requires first dividing the multidimensional vector space into grid cells and counting the number of mapped points (i.e., point density) within each grid cell. Then, the point density of each grid cell is divided by the maximum point density of all grid cells to obtain the relative density. For example, if a grid cell has 8 mapped points and a maximum point density of 10, then the relative density of that cell is 0.8.

[0088] Furthermore, the empirical confidence level is directly calculated using the relative density value. The higher the relative density, the more frequently the distressed state vector in that region appears in iterative predictions, and the higher the corresponding empirical confidence level. For example, if the relative density of the grid cell containing a mildly distressed vector is 0.8, then the empirical confidence level of that vector is 0.8, reflecting its high empirical support in backward predictions.

[0089] Furthermore, based on the voting confidence assessment results, the subjective confidence level of each point in the subjective distress state vector distribution is calculated to avoid the overall confidence assessment being affected by some low-confidence integrated prediction outputs.

[0090] The method provided in this application embodiment calculates the subjective confidence level of each point in the subjective distress state vector distribution based on the voting confidence assessment result, including: Based on a preset neighborhood radius, the neighborhood space of each point in the subjective distress state vector distribution is determined; Based on the voting confidence assessment results, obtain the set of voting confidence scores within the neighborhood space, calculate the mean of the set of voting confidence scores, and output it as the subjective confidence score. The subjective confidence level of each point is obtained through iterative calculation.

[0091] Specifically, the neighborhood space of each point in the subjective distress state vector distribution is first determined based on a preset neighborhood radius. The setting of the neighborhood radius needs to be combined with the density of the vector distribution and the accuracy requirements of labor monitoring. If the neighborhood radius is too small, the neighborhood space contains too few vector points, making it difficult to reflect the overall credibility level; if the neighborhood radius is too large, irrelevant vector points are easily included, leading to deviations in the confidence calculation.

[0092] For example, a mildly distressed vector has coordinates (3.2, 4.1, 2.3) in three-dimensional space. The neighborhood space defined by a radius of 0.4 will contain all vector points with coordinates in the range of (2.8-3.6, 3.7-4.5, 1.9-2.7). The distress states corresponding to these points are strongly correlated with the mildly distressed vector and can be used as a reference for confidence calculation.

[0093] Furthermore, based on the voting confidence assessment results, a set of voting confidence scores within the neighborhood space is obtained, and the mean of this set is calculated, outputting the subjective confidence score for that point. The voting confidence assessment results record the voting confidence score corresponding to each ensemble prediction output, while each point in the subjective distress state vector distribution corresponds to a distress state vector of a certain ensemble prediction output. Therefore, through vector matching, the voting confidence scores corresponding to all vector points within the neighborhood space can be extracted from the voting confidence assessment results, forming a confidence score set.

[0094] For example, the neighborhood space of the aforementioned mild distress vector contains 5 vector points with corresponding voting confidence levels of 0.46, 0.49, 0.47, 0.48, and 0.45, respectively. These values ​​are integrated to form a confidence set {0.46, 0.49, 0.47, 0.48, 0.45}.

[0095] Furthermore, the mean of this confidence set is calculated, i.e., (0.46+0.49+0.47+0.48+0.45) / 5=0.47. This mean is the subjective confidence score of the mildly distressed vector point. Through the above calculation method, the subjective confidence score incorporates both the overall credibility level of multiple vectors in the neighborhood and the predictive performance of the weak mapper, making it more valuable than empirical confidence scores that solely rely on vector density.

[0096] Furthermore, the subjective confidence level of each point is obtained through iterative calculation. That is, for each vector point in the distribution of subjective distress state vectors, the above steps of "determining the neighborhood space - extracting the confidence set - calculating the mean" are repeated to obtain the subjective confidence level of each point one by one.

[0097] Furthermore, the distribution of subjective distress state vectors and subjective confidence levels are correlated. Specifically, a correlation data table is first constructed, containing three columns: vector coordinates, distress state description, and subjective confidence level. The three-dimensional coordinates of each vector point, the corresponding distress state, and the calculated subjective confidence level are sequentially filled into the table to form structured correlation data, which is convenient for subsequent retrieval and reference.

[0098] Furthermore, to present the correlation results intuitively, it is also necessary to output the correspondence between the distribution of subjective distress state vectors and subjective confidence levels through visualization. For example, in a three-dimensional vector space graph, different colored points represent different distress states, and the size or color intensity of the points reflects the level of subjective confidence.

[0099] For example, taking a mildly distressed vector point as an example, if its subjective confidence level is 0.47, it can be set as a medium-sized yellow dot; if another mildly distressed vector point has a subjective confidence level of 0.6, it can be set as a larger dark yellow dot, so that the difference in confidence level can be directly reflected by visual difference.

[0100] In addition, when associating outputs, it is necessary to ensure the integrity and consistency of the data, that is, each vector point in the subjective distress state vector distribution has one and only one corresponding subjective confidence in the association result, with no omissions or duplicate annotations.

[0101] Ultimately, by linking structured tables with visual graphics, the system can provide accurate structured data support for subsequent calculations of monitoring abnormalities, offer clear reference for early warning decisions in labor monitoring, and further ensure the accuracy and practicality of early warnings.

[0102] S140: Calculate the monitoring abnormality degree based on the distribution of the real-time distress state vector and the subjective distress state vector, and provide early warning for labor monitoring based on the preset early warning threshold.

[0103] In this embodiment of the application, in order to avoid the problem of one-sided risk assessment and lack of accurate basis for early warning due to relying on a single state vector, it is necessary to calculate the monitoring abnormality degree through multiple steps and trigger the early warning in combination with a preset early warning threshold, so as to achieve accurate early warning and drug intervention prompts for fetal heart rate abnormalities and ensure the safety of mother and baby during delivery.

[0104] Specifically, an ensemble prediction output is first arbitrarily selected as the reference distress state vector set. The reference distress state vector set is derived from the ensemble prediction output obtained by the back-forward prediction of the second prediction model in the previous stage, and contains multiple different distress state vectors. When selecting, it is necessary to ensure that it can cover the main state types in the subjective distress state vector distribution.

[0105] Furthermore, historical fetal normal state vectors are obtained, and the corresponding global covariance matrix is ​​calculated. The historical fetal normal state vectors are derived from clinically confirmed normal delivery cases, covering the characteristics of normal fetal states at different gestational weeks and under different maternal health backgrounds. By collecting a sufficient number of historical normal state vectors, a representative normal state dataset can be constructed.

[0106] When calculating the global covariance matrix, it is necessary to quantify the correlation between different state dimensions based on the normal state dataset in order to eliminate the interference of the difference in data scale of different dimensions on the subsequent distance calculation and ensure the scientific nature of the anomaly calculation.

[0107] Furthermore, by combining the global covariance matrix, the Mahalanobis distance between the real-time distressed state vector and each reference distressed state vector in the reference distressed state vector group is calculated.

[0108] Mahalanobis distance, which corrects for correlations between variables through the global covariance matrix, is more suitable for analyzing multi-dimensional medical data than traditional Euclidean distance. Its calculation results can quantify the degree of difference between real-time distress states and each reference distress state. By calculating the Mahalanobis distance between the real-time vector and each reference vector, the position and abnormal tendency of the real-time state in the distribution of subjective distress state vectors can be determined.

[0109] Furthermore, using empirical confidence level as the first weight and subjective confidence level as the second weight, multiple Mahalanobis distances are weighted to obtain the monitoring abnormality score. The monitoring abnormality score obtained through weighted calculation can comprehensively quantify the degree of abnormality in the real-time fetal status, providing accurate numerical basis for early warning.

[0110] Finally, labor monitoring alerts are issued based on preset warning thresholds and medication information. These thresholds are set in conjunction with clinical standards for classifying fetal distress risk levels. When the calculated level of monitoring abnormality exceeds a certain threshold, the corresponding level of alert is automatically triggered, and previously acquired real-time medication information is simultaneously linked. By linking alerts with medication information, the alerts become more targeted, helping medical staff quickly identify intervention directions and improving the efficiency of emergency response in labor monitoring.

[0111] Step S140 in the method provided in this application embodiment includes: Arbitrarily select one of the integrated prediction outputs as the reference distress state vector set; Obtain the historical fetal normal state vector and calculate the corresponding global covariance matrix; Based on the global covariance matrix, calculate the Mahalanobis distance between the real-time distress state vector and each reference distress state vector in the reference distress state vector group; The monitoring anomaly degree is obtained by weighting multiple Mahalanobis distances using the empirical confidence level as the first weight and the subjective confidence level as the second weight.

[0112] In this embodiment of the application, in order to quantitatively compare the real-time fetal distress status with the distribution of potential distress risks, it is necessary to calculate the monitoring abnormality degree by integrating real-time vectors, historical normal data and confidence information through multiple steps, so as to achieve accurate quantification of the risk of fetal heart rate abnormality and provide judgment indicators for carrying out delivery monitoring and early warning based on drug administration information in combination with early warning thresholds.

[0113] Specifically, a set of reference distress state vectors is first randomly selected from the ensemble prediction output. The ensemble prediction output is derived from the back-prediction results of the second prediction model. Each ensemble prediction output contains multiple different distress state vectors. When selecting, it is necessary to ensure that the reference vector set can cover the main state types in the subjective distress state vector distribution, such as typical vectors such as no distress, mild distress, and moderate distress, to avoid the subsequent anomaly calculation failing to fully reflect the difference between real-time state and potential risk due to a single reference vector.

[0114] Furthermore, a historical fetal normal state vector is obtained, and the corresponding global covariance matrix is ​​calculated. This historical fetal normal state vector is derived from clinically confirmed normal delivery cases, which must cover different gestational weeks and maternal health backgrounds to ensure broad representativeness of the vector.

[0115] For example, multiple sets of normal state vectors containing data on normal fetal heart rate baseline, good uterine contraction tolerance, and no underlying health risks are collected to form a historical normal dataset.

[0116] When calculating the global covariance matrix, it is necessary to quantify the correlation between different state dimensions based on historical normal datasets. For example, the correlation between the fetal heart rate risk dimension and the uterine contraction tolerance dimension. If the fetal heart rate baseline is stable, the uterine contraction tolerance is generally high, and the two are negatively correlated. The covariance matrix can incorporate the above correlation into the calculation to eliminate the interference of differences in data scales of different dimensions on subsequent distance calculations and ensure the scientific validity of Mahalanobis distance calculations.

[0117] Furthermore, by combining the global covariance matrix, the Mahalanobis distance between the real-time distress state vector and each reference distress state vector in the reference distress state vector group is calculated. The Mahalanobis distance, unlike the traditional Euclidean distance, can correct for correlations between variables through the global covariance matrix, better meeting the analytical needs of multi-dimensional medical data.

[0118] In the specific calculation, the real-time distress state vector is paired one by one with each vector in the reference vector group, and the Mahalanobis distance between them is calculated by substituting them into the global covariance matrix. The larger the distance value, the more significant the difference between the real-time state and the reference state, and the higher the potential anomaly risk; the smaller the distance value, the closer the real-time state is to the reference state, and the relatively lower the risk. By calculating one by one, the anomaly tendency of the real-time state in the distribution of the subjective distress state vector can be fully grasped.

[0119] For example, if the real-time distress state vector is (fetal heart rate risk 3.1, uterine contraction tolerance 4.2, baseline health risk 2.2), the reference distress state vector group contains three vectors: reference vector 1 (fetal heart rate risk 1.0, uterine contraction tolerance 2.1, baseline health risk 1.1), reference vector 2 (fetal heart rate risk 2.4, uterine contraction tolerance 3.2, baseline health risk 1.7), and reference vector 3 (fetal heart rate risk 3.5, uterine contraction tolerance 4.5, baseline health risk 2.5). The global covariance matrix has been calculated, in which the covariance between fetal heart rate risk and uterine contraction tolerance is -0.3 (reflecting a negative correlation between the two, i.e., low fetal heart rate risk corresponds to high uterine contraction tolerance), the covariance between fetal heart rate risk and baseline health risk is 0.2, and the covariance between uterine contraction tolerance and baseline health risk is 0.15.

[0120] Substituting the real-time distress state vector and reference vector 1 into the global covariance matrix, the Mahalanobis distance is calculated to be 2.8. This relatively large distance indicates a significant difference between the real-time state and the low-risk reference vector 1, suggesting a higher potential anomaly risk. Calculating the distance between the real-time vector and reference vector 2, the Mahalanobis distance is 1.3, indicating a moderate difference between the real-time state and the medium-to-low-risk reference vector 2, with the risk falling within the transitional range. Calculating the distance between the real-time vector and reference vector 3, the Mahalanobis distance is 0.6, reflecting a smaller difference between the real-time state and the medium-risk reference vector 3, suggesting a risk tendency closer to the reference state.

[0121] Furthermore, multiple Mahalanobis distances were weighted with empirical confidence as the first weight and subjective confidence as the second weight to obtain the degree of abnormality in monitoring.

[0122] Specifically, in the weighted calculation, each Mahalanobis distance is multiplied by its corresponding empirical confidence level and subjective confidence level, summed, and then divided by the sum of the two confidence levels (to ensure weight normalization). For example, if a Mahalanobis distance is 2.5, corresponding to an empirical confidence level of 0.6 and a subjective confidence level of 0.48, and another Mahalanobis distance is 1.8, corresponding to an empirical confidence level of 0.55 and a subjective confidence level of 0.5, then the weighted calculation is (2.5 × 0.6 × 0.48 + 1.8 × 0.55 × 0.5) / (0.6 + 0.48 + 0.55 + 0.5). The final value obtained is the degree of abnormality in monitoring.

[0123] Ultimately, by using the weighted calculation method described above, the Mahalanobis distance corresponding to the reference vector with high confidence level can account for a higher proportion in the abnormality calculation, reducing the interference of low confidence vectors on the results and ensuring that the monitoring abnormality level can comprehensively quantify the degree of abnormality in the real-time fetal status.

[0124] Furthermore, after obtaining the abnormality level of monitoring, delivery monitoring and early warning based on the preset early warning threshold are carried out.

[0125] The preset warning thresholds need to be determined by combining the fetal distress risk level classification standards in clinical practice with a large amount of historical case data, so as to ensure that the thresholds can accurately distinguish different risk levels and meet the emergency response needs in the context of labor monitoring.

[0126] For example, the monitoring abnormality is divided into three intervals: low risk (<1.0), medium risk (1.0-2.0), and high risk (>2.0). Two warning thresholds of 1.0 and 2.0 are set accordingly. No warning is triggered in the low risk interval, a yellow warning is triggered in the medium risk interval, and a red warning is triggered in the high risk interval. Different warning levels correspond to different intervention prompt strategies.

[0127] At the same time, the early warning process needs to be closely linked to the real-time drug administration information obtained in the early stage, so that the early warning results can directly point to the direction of drug administration adjustment and avoid the disconnect between the early warning and actual medical intervention.

[0128] For example, if the calculated monitoring abnormality level is 2.3, which exceeds the high-risk threshold of 2.0, a red alert is triggered. If the real-time drug administration information shows that the mother is using oxytocin and the current drip rate is 8 units / hour, the alert device will simultaneously prompt "the monitoring abnormality level has reached a high risk. It is recommended to immediately stop the oxytocin infusion and recheck the fetal heart rate and uterine contractions."

[0129] In addition, if the monitoring abnormality level is 1.5, which is in the medium-risk range, a yellow alert will be triggered, and analgesics will be administered in real time. The alert will indicate that "the monitoring abnormality level has reached the medium-risk level, and it is necessary to closely monitor changes in fetal heart rate and assess whether the dosage of analgesics needs to be adjusted."

[0130] Ultimately, by using the aforementioned threshold-based warning system to correlate with medication administration, we can ensure the accuracy and timeliness of warnings while providing clear intervention guidelines for medical staff. This avoids the problem of traditional warnings that only indicate risks without specific response directions, effectively improving the efficiency of handling fetal heart rate abnormalities during labor monitoring and ensuring maternal and infant safety.

[0131] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects: This application proposes an intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities. First, it acquires the target mother's labor monitoring information, medical records, and real-time medication information. Next, based on the labor monitoring and medical records, a first prediction model is used to forward predict the fetal distress state, obtaining a real-time distress state vector. Then, a second prediction model is constructed, and the medical records and real-time medication information are input to backward predict the fetal distress state, obtaining a subjective distress state vector distribution. Next, the medical records and medication information are input into the second prediction model for iterative prediction, constructing a multi-dimensional vector space to map the distress state vector distribution. The relative density distribution is calculated to define the empirical confidence level, and the subjective confidence level is calculated based on the voting confidence assessment results. The output distribution and confidence level are correlated to cover potential risks and credibility. Finally, the monitoring abnormality degree is calculated based on the real-time distress state vector and the subjective distress state vector distribution. Combined with a preset early warning threshold, a labor monitoring early warning based on medication information is provided, achieving comprehensive perception, accurate quantification, and targeted early warning of fetal heart rate abnormality risk during labor monitoring.

[0132] The method provided in this application, through the technical solution of "multi-source information acquisition - positive prediction of real-time status - reverse deduction of risk distribution - quantitative abnormality warning", solves the problems of traditional labor monitoring relying on manual interpretation, partial information coverage leading to delayed warnings, and inaccurate risk assessment. It improves the accuracy and timeliness of fetal heart rate abnormality warnings and provides reliable technical support for obstetric clinical monitoring.

[0133] Example 2, as shown in the appendix Figure 2 As shown, based on the inventive concept of the intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities provided in Embodiment 1, this application also provides an intelligent early warning system for labor monitoring in response to fetal heart rate abnormalities, specifically including: The maternal multi-source information acquisition module 01 is used to acquire the target maternal labor monitoring information, medical record information and real-time drug administration information, wherein the labor monitoring information includes at least fetal heart rate timing data and uterine contraction timing data; The real-time distress positive prediction module 02 is used to predict the fetal distress state positively through the first prediction model based on the labor monitoring information and the medical record information, and obtain the real-time distress state vector. The second model reverse prediction module 03 is used to construct a second prediction model and input the medical record information and the real-time drug administration information into the second prediction model to reverse predict the fetal distress state and obtain the subjective distress state vector distribution. The monitoring anomaly calculation and early warning module 04 is used to calculate the monitoring anomaly degree based on the distribution of the real-time distress state vector and the subjective distress state vector, and to provide early warning for labor monitoring based on the preset early warning threshold.

[0134] In one embodiment, the second model inverse prediction module 03 is further configured to: A reverse analysis indicator set is determined, and samples are collected based on the reverse analysis indicator set to obtain reverse analysis sample data. The reverse analysis sample data includes at least a sample reverse indicator vector and a corresponding sample distress state vector. Using the reverse analysis sample data as training data, multiple machine learning-based weak mappers are constructed and trained. The weak mappers use the sample reverse indicator vector as the mapping input and the sample distress state vector as the mapping supervision. The multiple weak mappers are ensembled using an ensemble learning method to obtain an ensemble prediction core. The ensemble prediction core is then connected to a voting confidence evaluation layer to obtain the second prediction model. The medical record information and the real-time drug administration information are input into the second prediction model, and a preset number of iterative predictions are performed to obtain multiple integrated prediction outputs and corresponding voting confidence assessment results. A multidimensional vector space is constructed, and the distress state vectors in the multiple integrated prediction outputs are mapped to the multidimensional vector space to obtain the subjective distress state vector distribution. The relative density distribution of the subjective distress state vector distribution is calculated, and the empirical confidence level of each point in the subjective distress state vector distribution is defined according to the relative density distribution. Based on the voting confidence assessment results, the subjective confidence level of each point in the subjective distress state vector distribution is calculated accordingly. The subjective distress state vector distribution and the subjective confidence level are then correlated and output.

[0135] Furthermore, the second model inverse prediction module 03 also includes: Based on prior knowledge of fetal distress, relevant indicators are obtained, and a set of correlation coefficients between the relevant indicators and fetal distress is obtained through association analysis. Based on the medical record information and the real-time medication information, current status indicators are extracted, and the intersection of the current status indicators and the relevant indicators is taken. The set of correlation coefficients is normalized, and the cumulative correlation coefficient corresponding to the intersection is calculated. If the cumulative correlation coefficient is less than a preset cumulative limit, the intersection is output as the reverse analysis indicators; otherwise, the intersection is serialized, and the first N intersection factors that satisfy the cumulative limit are selected and output as the reverse analysis indicators.

[0136] Furthermore, the second model inverse prediction module 03 also includes: Obtain the integrated prediction output of the integrated prediction core, wherein the integrated prediction output includes multiple distressed state vectors and corresponding voting ratios; obtain the confidence scores of multiple weak mappers corresponding to the distressed state vectors, and calculate the mean confidence score; calculate the product of the mean confidence score and the voting ratio, and output the voting confidence score; iterate through and obtain the voting confidence scores corresponding to multiple distressed state vectors, and output the voting confidence evaluation result.

[0137] Furthermore, the second model inverse prediction module 03 also includes: Based on a preset neighborhood radius, the neighborhood space of each point in the subjective distress state vector distribution is determined; according to the voting confidence assessment result, the voting confidence set in the neighborhood space is obtained, and the mean of the voting confidence set is calculated and output as the subjective confidence; the subjective confidence of each point is obtained by iterative calculation.

[0138] In one embodiment, the monitoring anomaly calculation and early warning module 04 is also used for: Arbitrarily select one of the integrated prediction outputs as the reference distress state vector group; obtain the historical fetal normal state vector and calculate the corresponding global covariance matrix; combine the global covariance matrix to calculate the Mahalanobis distance between the real-time distress state vector and each reference distress state vector in the reference distress state vector group; use the empirical confidence level as the first weight and the subjective confidence level as the second weight to weight multiple Mahalanobis distances to obtain the monitoring abnormality degree.

[0139] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0140] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0141] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for intelligent early warning of labor monitoring in response to fetal heart rate abnormalities, characterized in that, include: Acquire the labor monitoring information, medical record information, and real-time drug administration information of the target pregnant woman, wherein the labor monitoring information includes at least fetal heart rate timing data and uterine contraction timing data; Based on the labor monitoring information and the medical record information, the fetal distress state is positively predicted through the first prediction model to obtain a real-time distress state vector; A second prediction model is constructed, and the medical record information and the real-time drug administration information are input into the second prediction model to predict the fetal distress state in reverse and obtain the subjective distress state vector distribution. Based on the distribution of the real-time distress state vector and the subjective distress state vector, the monitoring abnormality degree is calculated, and a labor monitoring early warning based on the preset early warning threshold is performed.

2. The intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities as described in claim 1, characterized in that, Constructing a second prediction model includes: Determine the reverse analysis indicators, and collect samples based on the reverse analysis indicators to obtain reverse analysis sample data, wherein the reverse analysis sample data includes at least the sample reverse indicator vector and the corresponding sample distress state vector. Using the reverse analysis sample data as training data, multiple machine learning-based weak mappers are constructed and trained, wherein the weak mappers use the sample reverse indicator vector as the mapping input and the sample distress state vector as the mapping supervision. The multiple weak mappers are integrated by voting using an ensemble learning method to obtain an ensemble prediction core, and the ensemble prediction core is connected to a voting confidence evaluation layer to obtain the second prediction model.

3. The intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities as described in claim 2, characterized in that, Determine the reverse analysis indicators, and collect samples based on the reverse analysis indicators to obtain reverse analysis sample data, including: Based on prior knowledge of fetal distress, relevant indicators are obtained, and through association analysis, a set of association coefficients between the relevant indicators and fetal distress is obtained. Based on the medical record information and the real-time drug administration information, extract the current status indicators and take the intersection of the current status indicators and the relevant indicators; Normalize the set of correlation coefficients and calculate the cumulative correlation coefficients corresponding to the intersection; If the cumulative correlation coefficient is less than the preset cumulative limit, the intersection is output as the reverse analysis index; otherwise, the intersection is serialized and the first N intersection factors that satisfy the cumulative limit are selected and output as the reverse analysis index.

4. The intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities as described in claim 3, characterized in that, The execution steps of the voting confidence assessment layer include: Obtain the integrated prediction output of the integrated prediction core, wherein the integrated prediction output includes multiple distress state vectors and corresponding voting ratios; Obtain the confidence scores of multiple weak mappers corresponding to the distress state vector, calculate the mean confidence score, calculate the product of the mean confidence score and the voting ratio, and output the voting confidence score. The voting confidence scores corresponding to multiple distress state vectors are obtained through iteration, and the output is the voting confidence evaluation result.

5. The intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities as described in claim 4, characterized in that, Input the medical record information and the real-time drug administration information into the second prediction model to predict the fetal distress state in reverse, and obtain the subjective distress state vector distribution, including: Input the medical record information and the real-time drug administration information into the second prediction model, perform a preset number of iterative predictions, and obtain multiple integrated prediction outputs and corresponding voting confidence assessment results; Construct a multidimensional vector space and map the distress state vectors from multiple integrated prediction outputs to the multidimensional vector space to obtain the distribution of the subjective distress state vectors. Calculate the relative density distribution of the subjective distress state vector distribution, and define the empirical confidence level of each point in the subjective distress state vector distribution based on the relative density distribution; Based on the voting confidence assessment results, the subjective confidence level of each point in the subjective distress state vector distribution is calculated accordingly; The subjective distress state vector distribution is correlated with the subjective confidence level.

6. The intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities as described in claim 5, characterized in that, Based on the voting confidence assessment results, the subjective confidence level of each point in the subjective distress state vector distribution is calculated, including: Based on a preset neighborhood radius, the neighborhood space of each point in the subjective distress state vector distribution is determined; Based on the voting confidence assessment results, obtain the set of voting confidence scores within the neighborhood space, calculate the mean of the set of voting confidence scores, and output it as the subjective confidence score. The subjective confidence level of each point is obtained through iterative calculation.

7. The intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities as described in claim 6, characterized in that, Based on the distributions of the real-time distress state vector and the subjective distress state vector, the degree of monitoring abnormality is calculated, including: Arbitrarily select one of the integrated prediction outputs as the reference distress state vector set; Obtain the historical fetal normal state vector and calculate the corresponding global covariance matrix; Based on the global covariance matrix, calculate the Mahalanobis distance between the real-time distress state vector and each reference distress state vector in the reference distress state vector group; The monitoring anomaly degree is obtained by weighting multiple Mahalanobis distances using the empirical confidence level as the first weight and the subjective confidence level as the second weight.

8. A smart early warning system for labor monitoring in response to fetal heart rate abnormalities, characterized in that: The system is used to execute the intelligent early warning method for labor monitoring in response to fetal heart rate abnormalities as described in any one of claims 1-7, the system comprising: The maternal multi-source information acquisition module is used to acquire the target maternal labor monitoring information, medical record information, and real-time drug administration information, wherein the labor monitoring information includes at least fetal heart rate timing data and uterine contraction timing data; The real-time distress positive prediction module is used to predict the fetal distress state positively through the first prediction model based on the labor monitoring information and the medical record information, and obtain the real-time distress state vector. The second model reverse prediction module is used to construct a second prediction model and input the medical record information and the real-time drug administration information into the second prediction model to reverse predict the fetal distress state and obtain the subjective distress state vector distribution. The monitoring anomaly calculation and early warning module is used to calculate the monitoring anomaly degree based on the distribution of the real-time distress state vector and the subjective distress state vector, and to provide early warning for labor monitoring based on the preset early warning threshold.

Images