An AI-based labor progress monitoring system

Abstract

The present application belongs to the technical field of intelligent monitoring of labor process, and discloses an AI-based labor process progress monitoring system, which comprises a labor process multi-source sensing module, a labor process time sequence state construction module, a labor process stage identification module, a trajectory constraint reinforcement learning module, a progress deviation checking module and a monitoring feedback output module; labor process state reliable fusion vectors are generated according to data sources, data timeliness and signal quality, labor process stages are identified by using causal time sequence coding, target monitoring actions are generated by double progress benefit evaluation, safety risk evaluation, safety action shielding and delayed outcome rewriting, and labor process progress deviation states are generated in combination with cervical dilation, fetal head descent and uterine contraction response deviation; the labor process review time can be dynamically adjusted, the influence of interference data and overestimation of action value is reduced, and the timeliness, safety and traceability of labor process monitoring are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent labor monitoring technology, specifically relating to an AI-based labor progress monitoring system. Background Technology

[0002] Labor is characterized by significant variations in duration, unstable transition speeds between stages, and diverse sources of monitoring data. Obstetric medical staff need to comprehensively assess the progress of labor by considering fetal heart rate, uterine contractions, maternal vital signs, cervical dilation, fetal head descent, and bedside ultrasound results. They also need to adjust the review time and monitoring density in a timely manner based on changes in the labor stages. Therefore, it is necessary to establish an intelligent monitoring system that can integrate continuous monitoring data and intermittent examination data, predict changes in the labor process, and assist in generating review tasks.

[0003] Current labor monitoring mainly involves cervical examinations at fixed time intervals, combined with fetal heart rate and uterine contraction monitoring results for manual judgment. Data from different devices and manual examinations lack a unified timeline and reliability evaluation. Early cervical examination results may continue to affect the current judgment even after a long period of non-review. Probe detachment, signal mutations, and inconsistent examination results can also be directly input into the recognition model, resulting in the labor stage recognition results not accurately reflecting the current true progress.

[0004] Furthermore, existing AI labor monitoring systems typically use fixed classification models to output labor stages and risk scores, making it difficult to dynamically determine the review interval based on the different progress rates of different mothers. Some reinforcement learning schemes mainly select monitoring actions based on short-term prediction accuracy, lacking constraints on the order of labor stages, safety risk assessment, and high-risk action shielding. At the same time, the delayed results from subsequent cervical examinations and fetal head descent examinations cannot be retrospectively evaluated to assess previous monitoring decisions, which can easily lead to overestimation of action value, delayed review prompts, and unsafe exploration issues. Summary of the Invention

[0005] This invention aims to address the problems of existing labor monitoring methods that rely primarily on fixed examination intervals and human experience, lack unified temporal integration of continuous fetal heart rate and uterine contraction data, maternal vital signs data, and intermittent labor examination data, and whose examination results continue to affect labor progress assessment even after they become outdated over time. Furthermore, traditional AI models only output labor stages and risk scores, failing to dynamically determine the timing of review based on the different progress rates of different mothers. It also addresses the issues of conventional reinforcement learning, which focuses on short-term predictive gains as its primary optimization objective, easily leading to overestimation of value, unsafe exploration, and delays in clinical outcomes from being fed back to preceding monitoring actions.

[0006] To address the aforementioned technical problems, this invention provides an AI-based labor progress monitoring system, comprising a multi-source labor perception module, a labor time sequence state construction module, a labor stage identification module, a trajectory constraint reinforcement learning module, a progress deviation verification module, and a monitoring feedback output module.

[0007] The multi-source labor sensing module collects fetal heart rate data, uterine contraction data, maternal vital signs data, labor examination data, maternal basic data, and labor intervention status data, and saves data collection time, data source, signal quality, and examination personnel information to generate a multi-source labor monitoring dataset.

[0008] The labor process timeline status construction module receives the multi-source monitoring dataset of the labor process, divides the continuous monitoring data into monitoring segments, writes the labor process examination data as event-type data into a unified labor process timeline, and performs reliable correction on various types of data based on the reliability of the data source, the time interval from the most recent effective collection, the data missing status, and the data mutation status, and generates a reliable fusion vector of the labor process status.

[0009] The labor stage identification module receives a reliable fusion vector of labor status and uses a causal temporal coding structure to extract changes in cervical dilation, fetal head descent, uterine contraction effectiveness, fetal heart rate status, and data reliability. It then generates the probability of latent stage, active stage, second stage of labor, stage transition probability, and labor progress prediction trajectory.

[0010] The labor stage identification module sets stage sequence constraints. In the absence of valid labor examination results and abnormal data quality records, it restricts the occurrence of reverse jumps in labor stages that are inconsistent with the physiological process, so that the stage identification results conform to the temporal continuity of labor development.

[0011] The trajectory constraint reinforcement learning module receives the reliable fusion vector of labor status, the probability of labor stage, the probability of stage transition, and the predicted trajectory of labor progress. It selects target monitoring actions from the preset monitoring action set through the labor trajectory constraint backtracking reinforcement learning algorithm. The monitoring actions include maintaining the basic monitoring density, shortening the interval of labor review, increasing bedside ultrasound review tasks, generating medical and nursing review prompts, and generating high-level medical and nursing review prompts.

[0012] The labor trajectory constraint backtracking reinforcement learning algorithm includes a first progress benefit evaluator, a second progress benefit evaluator, a safety risk evaluator, a stage consistency reward unit, a safety action shielding unit, and a delayed outcome write-back unit. The first and second progress benefit evaluators evaluate the degree to which monitoring actions improve the timeliness of labor progress identification, and select the lower evaluation result to limit overestimation of value. The safety risk evaluator evaluates the risk of missed reports and the risk of delayed review. The stage consistency reward unit generates action rewards based on the consistency between the prediction stage and the subsequent confirmation stage, the amount of advance warning, the number of erroneous warnings, the additional examination burden, and deviations from the physiological trajectory.

[0013] The safety action shielding unit generates a safety action shielding matrix based on fetal heart rate status, maternal vital signs status, deviation from labor progress status, and data uncertainty. When the safety shielding conditions are met, actions that maintain the basic monitoring density are removed from the candidate action set, thus restricting reinforcement learning from selecting low monitoring density actions under high-risk conditions.

[0014] After subsequent cervical examinations, fetal head descent examinations, and bedside ultrasound examinations are completed, the delayed outcome write-back unit writes the actual examination results back to multiple labor stages within the risk retrospective window, enabling the monitoring actions that generate review tasks in advance to receive delayed evaluation, and using the write-back results to update the first and second progress benefit evaluators.

[0015] The progress deviation verification module receives the predicted trajectory of labor progress and subsequent actual examination results. Based on the cervical dilation prediction deviation, fetal head descent prediction deviation, uterine contraction response deviation, confidence level of labor stage identification, and data uncertainty, it generates a labor progress deviation value. Based on the labor progress deviation value and its continuous change status, it generates a stable progress status, a status pending review, and a status requiring key review.

[0016] The progress deviation verification module further identifies deviations in cervical progress, fetal head descent, uterine contraction response, and data quality, enabling the monitoring feedback output module to display the specific source of the labor deviation and avoid directly judging abnormal labor as probe detachment, expired examination data, or inconsistencies in multi-source data.

[0017] The monitoring feedback output module receives the labor stage identification results, target monitoring actions, and labor progress deviation status, and generates labor monitoring results including the current labor stage, stage duration, cervical dilation trend, fetal head descent trend, uterine contraction effectiveness trend, labor progress prediction interval, main sources of deviation, and suggested review time.

[0018] The monitoring feedback output module receives the acceptance status, execution time, actual examination results, and prompts from medical staff regarding the monitoring actions, generates monitoring feedback data, and sends the monitoring feedback data to the labor process time sequence status construction module, trajectory constraint reinforcement learning module, and progress deviation verification module, forming a closed-loop monitoring chain of multi-source perception, status construction, stage identification, monitoring action selection, deviation verification, and clinical result writing.

[0019] The system uses historical labor records to train the trajectory constraint reinforcement learning module offline, and performs shadow running verification before the formal output review prompt; the system output is limited to labor monitoring density adjustment and medical staff review tasks, and does not generate drug dosage, delivery method and treatment results. Clinical treatment is determined by medical staff based on the examination results.

[0020] The beneficial effects of this invention are as follows: (1) Based on the data source, data timeliness, data missing status and data mutation status, the continuous monitoring data and intermittent examination data are reliably integrated, and the labor stage is identified through the causal time sequence coding structure. This can reduce the continuous impact of expired examination results and equipment interference data on the judgment of labor process and solve the problem of inconsistent time and different reliability of multi-source labor process data.

[0021] (2) In the reinforcement learning structure, a dual progress benefit evaluator, a safety risk evaluator, a safety action shield and a phase consistency reward are set up. The dual evaluation results limit the overestimation of action value, and low monitoring density actions are shielded when the fetal heart rate, maternal status and data uncertainty reach the safety conditions, so as to solve the problem of lack of safety exploration boundary in the medical monitoring process of ordinary reinforcement learning.

[0022] (3) Using the results of subsequent real labor examinations, multiple monitoring states are written back, and combined with the cervical dilation prediction deviation, fetal head descent prediction deviation, uterine contraction response deviation and data quality risk, the labor progress deviation state is generated, so that the system can learn the continuous changes before the stage transition and dynamically adjust the timing of the review, thereby improving the timeliness, interpretability and traceability of labor progress monitoring. Attached Figure Description

[0023] The accompanying drawings are provided to further understand the present invention and form part of the specification. They are used together with the embodiments of the present invention to explain the invention and do not constitute a limitation thereof.

[0024] Figure 1 This is a flowchart of the labor progress monitoring system of the present invention. Detailed Implementation

[0025] Example 1, see Figure 1 The present invention provides an AI-based labor progress monitoring system, including a labor multi-source perception module, a labor time sequence state construction module, a labor stage identification module, a trajectory constraint reinforcement learning module, a progress deviation verification module, and a monitoring feedback output module.

[0026] This embodiment is applied to the labor monitoring scenario of full-term singleton cephalic presentation mothers in maternal and child health care institutions. The delivery room is equipped with 12 labor beds, each equipped with fetal heart rate and uterine contraction monitoring equipment, maternal vital signs monitoring equipment and bedside data terminal. Labor examination data is entered by medical staff through the bedside data terminal. The system only generates labor stage identification results, labor progress prediction results, monitoring interval adjustment results and review prompts, and does not directly control the equipment for induction, analgesia and delivery management.

[0027] The multi-source labor sensing module acquires fetal heart rate data, uterine contraction data, maternal vital signs data, labor examination data, maternal basic data, and labor intervention status data. Among them, fetal heart rate data includes fetal heart rate baseline, fetal heart rate variability, acceleration events, and deceleration events; uterine contraction data includes contraction frequency, contraction duration, and contraction intensity; maternal vital signs data includes heart rate, blood pressure, blood oxygen saturation, and body temperature; labor examination data includes cervical dilation, cervical effusion, fetal presenting position, fetal head descent, and fetal membrane status; maternal basic data includes gestational age, parity, age, and physical condition; and labor intervention status data includes analgesia status, intravenous infusion status, and artificial rupture of membranes records.

[0028] The labor process timeline construction module registers continuous monitoring data and intermittent examination data according to a unified timeline, divides fetal heart rate data and uterine contraction data into one-minute monitoring segments, divides maternal vital signs data into five-minute monitoring segments, writes labor examination data as event-type data into the corresponding time position, and generates a reliable fusion vector of labor process status based on the reliability of data source, data generation time, data missing status, and data mutation status.

[0029] The labor stage identification module receives a reliable fusion vector of labor status and uses a labor stage temporal encoder to extract changes in cervical dilation, fetal head descent, uterine contraction effectiveness, and fetal heart rate over the past sixty minutes. It then generates the probabilities of latent stage, active stage, second stage, and stage transition, and forms the labor stage identification result based on the stage transition probability.

[0030] The trajectory constraint reinforcement learning module receives the labor stage identification results and the labor status credible fusion vector. It determines the current monitoring action through the labor trajectory constraint backtracking reinforcement learning algorithm. The monitoring actions include maintaining the basic monitoring density, shortening the labor review interval, increasing bedside ultrasound review tasks, generating medical and nursing review prompts, and generating high-level review prompts. The reinforcement learning process does not output drug dosage or clinical treatment actions.

[0031] The trajectory-constrained reinforcement learning module includes a progress benefit evaluator, a safety risk evaluator, a phase consistency reward unit, a delayed outcome write-back unit, and a safety action shielding unit. The progress benefit evaluator predicts the degree to which monitoring actions improve the timeliness of labor progress identification. The safety risk evaluator predicts the underreporting risk and delayed review risk corresponding to the monitoring actions. The phase consistency reward unit restricts the reverse jump of labor stages when there is a lack of valid clinical evidence. The delayed outcome write-back unit writes the actual results back to the previous multiple monitoring states after the completion of subsequent cervical examinations and fetal head descent examinations. The safety action shielding unit prohibits reinforcement learning from performing low-monitoring-density actions when the fetal heart rate, maternal vital signs, and labor examination status meet the safety shielding conditions.

[0032] The progress deviation verification module receives the results of labor stage identification, the predicted labor trajectory generated by the trajectory constraint reinforcement learning module, and the subsequent actual examination results. Based on the cervical dilation prediction deviation, fetal head descent prediction deviation, uterine contraction response deviation, stage identification uncertainty, and data quality risk, it generates a labor progress deviation value and divides the labor progress deviation status into stable progress status, pending review status, and key review status.

[0033] The monitoring feedback output module generates a labor monitoring interface based on the labor stage identification results, monitoring actions, and deviations in labor progress. The labor monitoring interface displays the current labor stage, stage duration, cervical dilation trend, fetal head descent trend, uterine contraction effectiveness trend, progress prediction interval, main sources of deviation, and suggested review time. After confirmation by medical staff, the actual examination results are fed back to the labor time sequence status construction module and the trajectory constraint reinforcement learning module.

[0034] Through the above processing, this system constructs the labor status using continuous fetal heart rate and uterine contraction data and intermittent labor progress monitoring data. By improving the reinforcement learning algorithm, the monitoring density is dynamically determined under the condition of ensuring safe action boundaries. The system also uses subsequent real examination results to retrospectively correct the previous status evaluation, which solves the problems that fixed monitoring intervals cannot adapt to different labor progress speeds and that ordinary reinforcement learning tends to pursue short-term prediction accuracy while ignoring medical safety constraints.

[0035] Example 2: Based on all the above examples, the multi-source labor sensing module specifically includes a fetal heart rate and uterine contraction acquisition unit, a maternal status acquisition unit, a labor examination and recording unit, a bedside ultrasound access unit, and a unified time scale unit.

[0036] The fetal heart rate and uterine contraction acquisition unit continuously acquires fetal heart rate curves and uterine contraction pressure curves through fetal heart rate and uterine contraction monitoring equipment. Based on a one-minute processing window, it extracts the fetal heart rate baseline, short-term variability, number of accelerations, number of decelerations, number of uterine contractions, average duration of uterine contractions, and uterine contraction curve integrals, and retains the corresponding original curve index for medical staff to review.

[0037] The maternal status acquisition unit acquires the maternal heart rate and blood oxygen saturation every five minutes, non-invasive blood pressure every fifteen minutes, and acquires body temperature according to the set sampling cycle when the body temperature monitoring device is continuously connected. It also saves the device number, acquisition time and signal quality mark for each monitoring result.

[0038] The labor examination data entry unit is set up with a structured input interface. After completing the labor examination, medical staff enter the results of cervical dilation, cervical effusion, fetal presenting position, fetal position, fetal membrane status, and amniotic fluid observation. At the same time, the examiner's number and examination time are recorded to prevent the system from automatically inferring the results of manual examination as continuously changing data.

[0039] The bedside ultrasound access unit acquires the fetal head progression angle, fetal head-perineal distance, and fetal position results generated by ultrasound examination, and establishes a pairing relationship between the ultrasound measurement results and the results of the simultaneous manual examination. When the difference between the two types of results exceeds the verification range set by the institution, the manual examination results are not directly overwritten, but a data inconsistency mark is generated.

[0040] The unified time unit corrects the local time of each monitoring device based on the server clock, and maps continuous monitoring data, event-type inspection data and intervention status data to the labor time axis in a unified manner. When a short-term disconnection occurs in a continuous monitoring device, the missing data position is reserved, and the previous sampling value is not used to directly fill the entire disconnection period.

[0041] The multi-source perception module for labor further sets the source quality level for each data item. Labor examination data and bedside ultrasound data confirmed by medical staff are set as high-confidence event data, fetal heart rate and uterine contraction data with normal signal quality and maternal vital signs data are set as continuous and reliable data, and data with probe detachment, signal saturation and motion interference are set as low-confidence data.

[0042] When the fetal heart rate curve is briefly lost and the uterine contraction curve is normal, the system only reduces the reliability of fetal heart rate-related features and does not mark the entire labor process as invalid. When the uterine contraction probe is displaced and the baseline of the uterine contraction curve drifts, the system suspends the use of uterine contraction intensity to evaluate the progress status, but still retains the uterine contraction frequency and manually recorded information.

[0043] After labor progress data exceeds the set valid time, the system reduces the influence of cervical dilation and fetal head descent on the current state vector over time, and increases the candidate priority of the next labor progress review action, thereby preventing the long-term use of early examination results to form incorrect progress judgments.

[0044] Through the above processing, the system maps labor information with different sampling frequencies, data reliability, and data validity periods to a unified time axis, solving the problem of inconsistent status time caused by directly splicing continuous monitoring data and intermittent manual inspection data.

[0045] Example 3: Based on all the above examples, the labor process time-series status construction module specifically performs data quality assessment, time decay correction, short-term trend extraction, and reliable status fusion.

[0046] For the acquisition results of the k-th type of labor monitoring data at time position t, the labor time series status construction module reads the basic reliability of the data source, the time interval from the most recent valid acquisition, the data missing marker and the data mutation marker, and calculates the data reliability coefficient based on the above information.

[0047] Calculate the reliability coefficient and reliability correction result of the k-th type of labor monitoring data according to Formula 1:

[0048] In Formula 1, This represents the confidence coefficient of the k-th type of labor monitoring data at time position t. This indicates the basic reliability of the data source corresponding to the k-th type of labor monitoring data. This represents the time interval between the current time and the most recent valid data collection time. This represents the effective time scale corresponding to the k-th type of labor monitoring data. This indicates a missing data flag; a value of zero is used when data is valid, and a value close to one is used when data is missing. This indicates a data mutation marker. A zero value is assigned when the data change conforms to physiological continuity, and a higher value is assigned when the data exhibits device interference characteristics. This indicates the monitoring results after reliable correction. This indicates the current actual data collection results. This indicates a provisional estimate based on the most recent valid results and short-term trends.

[0049] The labor timeline status construction module does not use provisional estimates to replace clinical examination results. Provisional estimates only reflect the risk of missing data when reinforcement learning selects the next monitoring action, and the corresponding data is displayed as a trend estimate in the labor monitoring interface.

[0050] For cervical dilation and fetal head descent, the system generates the direction and rate of change based on the two most recent valid examination results, and sets confidence decay for the rate of change that lacks subsequent examination verification. For uterine contraction data, the system extracts the changes in uterine contraction frequency, duration, and effectiveness at three time scales: the most recent ten minutes, thirty minutes, and sixty minutes.

[0051] The labor process timeline state construction module combines the trusted corrected fetal heart rate status features, uterine contraction status features, maternal vital signs features, cervical status features, fetal head descent features, maternal basic characteristics, intervention status features, and various data trust coefficients into a trusted fusion vector of labor process status.

[0052] The labor stage temporal encoder uses a causal temporal convolutional structure to process the reliable fusion vector of labor status. The causal temporal convolutional structure only uses the current time and the data before the current time, and does not read subsequent inspection data to generate the current stage identification result, thereby ensuring that no future information is leaked during the online monitoring process.

[0053] The labor stage identification module further sets stage sequence constraints. After the labor stage moves from the latent stage to the active stage, the model is not allowed to return to the latent stage due to a single monitoring window change if there is a lack of medical staff confirmation records or obvious data quality abnormalities. After the second labor stage identification result is generated, it needs to be confirmed in combination with the cervical examination results and the fetal head descent results.

[0054] Regarding parameter adjustments: When the signal from the continuous monitoring equipment is stable, increase the basic reliability of the corresponding data source; when the interval between manual examinations is long, shorten the effective time scale of cervical status data; when the mother's position changes significantly and the external uterine contraction probe is re-fixed, re-establish the baseline of the uterine contraction curve.

[0055] Through the above processing, the system can construct the labor status based on data source, data timeliness, and signal quality, instead of inputting all monitoring data into the AI ​​model with the same weight, thus solving the problem of outdated intermittent examination results and equipment interference data continuously affecting the judgment of labor stage.

[0056] Example 4: Based on all the above examples, the trajectory constraint reinforcement learning module uses the labor trajectory constraint backtracking reinforcement learning algorithm to determine the monitoring action. The labor trajectory constraint backtracking reinforcement learning algorithm adds stage consistency reward, safety risk assessment, safety action masking and delayed outcome writing to the dual evaluator reinforcement learning structure.

[0057] The reinforcement learning status includes the probability of labor stage, the trend of cervical dilation, the trend of fetal head descent, the change in the effectiveness of uterine contractions, the fetal heart rate status, the status of maternal vital signs, the data reliability coefficient, the current monitoring interval, the time since the most recent labor examination, and the results of the most recent monitoring actions.

[0058] The reinforcement learning actions do not include clinical treatment procedures. The action set is limited to maintaining the current monitoring density, shortening the interval between manual reviews, increasing bedside ultrasound reviews, generating routine medical and nursing review prompts, and generating high-level medical and nursing review prompts. Medical staff can accept, postpone, or refuse monitoring actions and save the operation results as feedback data.

[0059] The phase consistency reward unit generates reinforcement learning rewards based on the degree of consistency between the prediction phase and the subsequent confirmation phase, the prediction error of labor progress, the amount of abnormal alerts, the number of false alerts, the additional examination burden, and the degree of inconsistency in physiological trajectory.

[0060] Calculate the labor monitoring action report corresponding to time position t according to Formula 2:

[0061] In Formula 2, This indicates the labor monitoring action report corresponding to time position t. This indicates the progress gain resulting from the improved accuracy of labor progress prediction due to the current action. This indicates the early recognition benefit generated before the actual deviation from the labor process is confirmed. This represents the cost of displaying error messages and repeated prompts. This indicates an increased monitoring burden due to manual examinations and bedside ultrasound examinations. This indicates a trajectory penalty for the predicted trajectory violating the sequence of labor stages and the continuity of physiological changes. This indicates the penalty for missed reporting due to failure to generate review actions in a timely manner. to These represent the return weights corresponding to each return item.

[0062] The progress benefit evaluator uses two independently initialized value networks to evaluate the same monitoring action and selects the lower prediction result of the two value networks as the progress benefit evaluation result, thereby reducing the probability that a single value network will overestimate the action of shortening the monitoring interval.

[0063] The safety risk evaluator predicts the delayed review risk corresponding to each monitoring action separately. When the fetal heart rate, maternal vital signs, and labor progress deviation meet the safety action shielding conditions, the safety action shielding unit removes the action that maintains the current monitoring density from the candidate action set, while retaining the medical staff review prompt action.

[0064] Perform reinforcement learning action selection and delayed outcome write-back update according to Formula 3:

[0065] In Formula 3, This indicates the target monitoring action selected at time position t. This represents the set of all monitored actions. This represents the safety action shielding matrix. This indicates that allowed actions are retained based on the security action masking matrix. These represent the first progress benefit evaluator and the second progress benefit evaluator, respectively. This indicates a safety risk assessor. This represents the safety risk mitigation coefficient. This indicates the uncertainty of the combination of states and actions. Indicates the uncertainty exploration coefficient. This indicates the goal of delayed outcome write-back, and N represents the number of monitoring states included in the risk backtracking window. This represents the time discount factor. This represents the target progress benefit evaluator. Indicates the allowed actions corresponding to the next state. This indicates the step size for value network updates.

[0066] Unlike ordinary hourly reward reinforcement learning, the labor trajectory constraint backtracking reinforcement learning algorithm, after subsequent manual checks confirm the degree of cervical dilation and the state of fetal head descent, the delayed outcome write-back unit writes the confirmation results back to multiple monitoring states in the risk backtracking window, so that the earlier generated actions that shorten the review interval can obtain the corresponding early identification benefits.

[0067] When the amount of training data is small, the system first uses the institution's historical labor records for offline training. During the offline training phase, it only learns from the monitoring actions that have actually been performed in the past and does not generate clinical treatment actions that do not appear in the historical data. After the system goes online, it first outputs monitoring suggestions in shadow mode but does not directly enter the medical and nursing workflow. The prompt function is only enabled after a set number of review verifications are completed.

[0068] Regarding parameter adjustments: In the initial stage of model deployment, the safety risk suppression coefficient is increased and the uncertainty exploration coefficient is reduced. After accumulating stable feedback data, the adaptive selection range between maintaining the current monitoring density and shortening the review interval is gradually expanded. When a missed event occurs, the weight of the missed event penalty is increased and the risk backtracking window is expanded.

[0069] Through the above processing, the reinforcement learning algorithm no longer selects actions solely based on the prediction accuracy of the next time window. Instead, it uses subsequent real labor examination results to retrospectively evaluate the previous monitoring decisions. At the same time, by using dual evaluators, safety risk assessment, and action masking to limit unsafe exploration, it solves the problem that ordinary reinforcement learning in medical monitoring scenarios is prone to overestimating value and lacking safety boundaries for exploration actions.

[0070] Example 5: Based on all the above examples, the progress deviation verification module specifically verifies the predicted labor trajectory and the actual labor status, and generates the labor progress deviation status according to different deviation sources.

[0071] The trajectory constraint reinforcement learning module predicts the degree of cervical dilation, fetal head descent, uterine contraction effectiveness, and probability of labor stage for the next review window after each effective labor check, and generates a prediction interval for each prediction result.

[0072] After subsequent labor progress checks are completed, the progress deviation verification module matches the actual degree of cervical dilation and the actual fetal head descent with the corresponding prediction results, and at the same time obtains the current uterine contraction effectiveness, the confidence level of labor stage identification, and data uncertainty.

[0073] Calculate the deviation of labor progress at time position t according to Formula 4:

[0074] In Formula 4, This represents the deviation of labor progress at time position t. Indicates the actual degree of cervical dilation. This indicates a prediction of the degree of cervical dilation. This indicates the actual descent of the fetal head. This indicates a prediction of fetal head descent. Indicates the effectiveness of actual uterine contractions. Indicates the effectiveness of predicting uterine contractions. This represents the normalized baseline for cervical dilation deviation. This represents the normalized baseline for the deviation of the fetal head descent. This represents the normalized baseline for the deviation of uterine contraction effectiveness. This indicates the confidence level for identifying the current stage of labor. This refers to data uncertainty caused by missing data, outdated data, and inconsistencies between multiple sources. These represent the evaluation weights corresponding to each deviation item.

[0075] When the cervical dilation prediction deviation increases while the fetal head descent prediction deviation remains low, the system generates a cervical progress deviation marker. When the fetal head descent prediction deviation increases while the uterine contraction effectiveness remains stable, the system generates a fetal head descent verification marker. When multiple prediction deviations are not high but data uncertainty continues to increase, the system generates a data supplement marker instead of directly generating a conclusion of abnormal labor.

[0076] The progress deviation verification module determines the deviation level based on the continuous change of the deviation value of the labor progress. When a single monitoring window reaches the deviation condition but the next window recovers, a short-term deviation record is retained. When multiple consecutive monitoring windows reach the deviation condition, the verification level is increased to avoid a single manual measurement difference directly triggering a high-level prompt.

[0077] When the deviation value of labor progress reaches the range to be reviewed, the trajectory constraint reinforcement learning module prioritizes shortening the labor progress review interval and increasing bedside ultrasound review tasks. When the deviation value of labor progress reaches the key review range and the safety risk assessor outputs a high risk, the system generates a high-level medical and nursing review prompt.

[0078] Regarding parameter adjustments: separate deviation evaluation parameter versions were established for primiparous and multiparous women. When bedside ultrasound data remained consistently stable, the evaluation weight of fetal head descent deviation was increased. When there were significant observer differences in manual examination data, the evaluation weight of data uncertainty was increased, and the direct impact of single cervical dilation deviation on the alert level was reduced.

[0079] Through the above processing, the system can distinguish between deviations from actual progress and data quality degradation, enabling review prompts to correspond to specific sources such as cervical progress, fetal head descent, uterine contraction response, and data inconsistencies, thus solving the problem of generating general alarms based solely on a single labor stage time threshold.

[0080] Example 6: Based on all the above examples, the monitoring feedback output module specifically includes a labor trajectory display unit, a review task generation unit, a medical feedback receiving unit, and a model version management unit.

[0081] The labor trajectory display unit displays the degree of cervical dilation, fetal head descent, uterine contraction effectiveness, and probability of labor stages according to a unified labor timeline. Actual examination results are displayed with solid markers, model trend estimation results are displayed with dashed lines, and prediction intervals are displayed with interval bands, enabling medical staff to distinguish between real examination data and AI prediction data.

[0082] The review task generation unit generates task content based on the monitoring actions output by the trajectory constraint reinforcement learning module. The task content includes the target pregnant woman number, suggested review items, suggested completion time, main triggering reasons, current data missing status, and corresponding sources of deviation from the labor progress.

[0083] The medical staff feedback receiving unit records the medical staff's acceptance status, execution time, examination results, and prompt evaluation of the review task. When the medical staff believe that the prompt is caused by probe detachment or data entry error, they select data problem feedback. When the medical staff confirms that there is a change in the progress of labor, they select valid prompt feedback.

[0084] When medical staff refuse to review a task, the system saves the reason for the refusal and does not immediately evaluate the refusal as a negative reward. After the actual labor check is completed, the system determines whether the preceding monitoring actions were reasonable based on the actual results, thereby avoiding the model from simply learning the operating habits of medical staff and deviating from the actual labor process results.

[0085] The model version management unit stores the production process stage time encoder version, progress benefit evaluator version, safety risk evaluator version, and parameter adjustment records respectively. The new model version first performs offline playback using historical data, and then compares it in parallel with the current model in shadow running mode. After reaching the stable conditions set by the mechanism, the version switch is performed.

[0086] Each model update retains a summary of the input data, the model version, monitoring actions, medical staff feedback, and subsequent confirmation results, enabling the same prompt to be traced back to the corresponding data status and algorithm version.

[0087] The system is equipped with a manual priority control interface, allowing medical staff to manually increase the monitoring level at any time. After the monitoring level is manually increased, the safety action shielding unit will not allow the reinforcement learning module to automatically decrease the monitoring level within a set time, preventing the algorithm output from overriding the active review decisions of clinical staff.

[0088] The labor stage identification results and review prompts generated by this system serve as auxiliary information for medical staff monitoring and do not replace obstetric examinations and clinical treatment decisions. Decisions regarding delivery methods and treatment plans shall be made by medical staff in accordance with clinical guidelines.

[0089] Through the above processing, the system output not only includes the labor stage and warning level, but also displays the source of the warning, the data credibility status, and the model prediction range. It also uses subsequent examinations and medical staff feedback to form a traceable update chain, solving the problem that traditional AI labor monitoring models output a single risk score and are difficult to interpret the basis of the warnings.

[0090] Example 7: Based on all the above examples, this example illustrates the specific operation of the system using a trial operation scenario in the delivery room of a maternal and child health care institution. The following data is only used to illustrate the system's data processing and reinforcement learning action selection process.

[0091] The trial participant was a full-term, singleton, cephalic primiparous woman. When she entered the labor monitoring phase, her cervical dilation was 3cm, her cervical effusion was 70%, her fetal presenting position was in the set initial position, her fetal heart rate monitoring signal quality was normal, and she had three contractions in the last ten minutes. The system identified the current state as a candidate state for transitioning from the latent stage to the active stage.

[0092] The multi-source perception module for labor continuously collects fetal heart rate curves, uterine contraction curves, and maternal vital signs data. The labor process time sequence status construction module generates a reliable fusion vector of labor status every five minutes. The cervical examination data maintains high reliability in the first 30 minutes after the first labor examination. As the time since the examination increases, the reliability coefficient of the cervical examination data gradually decreases.

[0093] Ninety minutes after the start of monitoring, the frequency and duration of contractions continued to increase, and the labor phase sequence encoder generated a high probability of active phases. However, the time scale corresponding to the current parameter version had exceeded the time since the most recent cervical examination, and the data uncertainty increased simultaneously.

[0094] The standard fixed interval scheme still requires waiting for the originally scheduled review time. The labor trajectory constraint backtracking reinforcement learning algorithm selects to shorten the labor review interval from the allowed actions based on the increase in the probability of stage transition, the decrease in the reliability of cervical data, and the continuous change in the trend of uterine contractions, and generates a cervical status review task.

[0095] After the medical staff completed the review, they recorded the degree of cervical dilation as 5cm. The system confirmed that the labor process had entered the active stage. The delayed outcome write-back unit wrote the examination results back to the previous multiple monitoring states, so as to generate the action in advance, shorten the review interval, and obtain the benefit of early identification of stage transition.

[0096] During subsequent monitoring, the external uterine contraction probe shifted position, the intensity of the uterine contraction curve decreased rapidly but the frequency of uterine contractions remained continuous, and the fetal heart rate and maternal vital signs did not change synchronously. The labor process time sequence status construction module improved the mutation marker of the uterine contraction intensity data and reduced the corresponding confidence coefficient.

[0097] The progress deviation verification module found that the contraction prediction deviation increased, but the data uncertainty also increased simultaneously. Therefore, it generated supplementary labels for contraction data instead of generating a conclusion of labor stagnation. The trajectory constraint reinforcement learning module selected to re-fix the monitoring probe and select the corresponding ordinary medical and nursing review prompt.

[0098] After the probe was repositioned, the uterine contraction curve returned to normal. The previous deviation was identified as a problem with the equipment signal. This prompt was recorded as a valid data quality prompt and was not included in the error alarm cost as a prompt of abnormal labor.

[0099] After continued monitoring, the system predicted that the degree of cervical dilation and the state of fetal head descent should change accordingly in the next review window. However, subsequent examinations showed that the changes in cervical dilation were lower than the predicted range and the changes in fetal head descent were also lower than the predicted range. The deviation value of labor progress increased for several consecutive windows.

[0100] The trajectory constraint reinforcement learning module removes actions that maintain basic monitoring density under the constraint of the safety action masking matrix, selects to add bedside ultrasound review tasks and generate high-level medical and nursing review prompts. The system interface shows that the main sources of deviation are cervical dilation prediction deviation and fetal head descent prediction deviation.

[0101] During this trial run, the reinforcement learning module only changed the monitoring density and review tasks, and did not output medication or delivery management suggestions. Medical staff independently made subsequent clinical judgments based on the obstetric examination results.

[0102] In offline comparative playback, ordinary reinforcement learning models without delayed outcome rewriting tend to maintain a fixed monitoring density because shortening the review interval increases the monitoring burden within the current window and does not immediately provide progress confirmation benefits. However, by adopting the labor trajectory constraint backtracking reinforcement learning algorithm, subsequent real examination results can be used to evaluate the early review actions, enabling the model to gradually learn the combination of uterine contraction changes, data expiration, and stage probability changes before the stage transition.

[0103] In another set of offline replays, the reinforcement learning model without safety action masking still chose to maintain the current monitoring density under some high uncertainty states. After adding the safety risk evaluator and the safety action masking matrix, when the data uncertainty and the risk of fetal heart rate status increase at the same time, low monitoring density actions cannot enter the candidate action set.

[0104] Through the above trial operation, the system has formed a closed loop of "multi-source labor perception - reliable temporal state construction - labor stage identification - trajectory constraint reinforcement learning to select monitoring actions - labor progress deviation verification - medical staff review and result retrospective update". This allows reinforcement learning to dynamically adjust the monitoring density according to the changes in labor of different mothers within the medical safety boundary, and to continuously correct the monitoring strategy by delaying clinical outcomes.

Claims

1. An AI-based labor progress monitoring system, characterized in that, It includes a multi-source perception module for labor process, a labor process time sequence state construction module, a labor process stage identification module, a trajectory constraint reinforcement learning module, a progress deviation verification module, and a monitoring feedback output module; The multi-source labor sensing module collects fetal heart rate data, uterine contraction data, maternal vital signs data, labor examination data, maternal basic data, and labor intervention status data. It also configures the collection time, data source, and signal quality markers for the collected data, generates a multi-source labor monitoring dataset, and sends the multi-source labor monitoring dataset to the labor time sequence status construction module. The labor process timeline state construction module receives the multi-source labor process monitoring dataset, maps continuous monitoring data and event-based labor process examination data to a unified labor process timeline, performs reliable correction on various types of monitoring data according to the reliability of data sources, data collection time, data missing status and data mutation status, generates a reliable fusion vector of labor process status, and sends the reliable fusion vector of labor process status to the labor process stage identification module and the trajectory constraint reinforcement learning module. The labor stage identification module receives the labor status credibility fusion vector, extracts changes in cervical dilation, fetal head descent, uterine contraction effectiveness, fetal heart rate status, and data credibility, generates labor stage identification results, stage transition probabilities, and labor progress prediction trajectories, and sends the labor stage identification results, stage transition probabilities, and labor progress prediction trajectories to the trajectory constraint reinforcement learning module and the progress deviation verification module; The trajectory constraint reinforcement learning module receives the reliable fusion vector of the labor state, the labor stage identification result, the stage transition probability, and the labor progress prediction trajectory. It then uses the labor trajectory constraint backtracking reinforcement learning algorithm to evaluate the progress benefits, safety risks, and safety actions of candidate monitoring actions, generates target monitoring actions, and sends the target monitoring actions to the monitoring feedback output module. The progress deviation verification module receives the predicted trajectory of labor progress and subsequent actual examination results, generates a labor progress deviation status based on the cervical dilation prediction deviation, fetal head descent prediction deviation, uterine contraction response deviation, confidence level of labor stage identification and data uncertainty, and sends the labor progress deviation status to the trajectory constraint reinforcement learning module and the monitoring feedback output module. The monitoring feedback output module receives the labor stage identification results, the target monitoring actions, and the labor progress deviation status, generates labor monitoring results and review tasks, receives the medical and nursing feedback and actual examination results corresponding to the review tasks, generates monitoring feedback data, and feeds the monitoring feedback data back to the labor time sequence status construction module, the trajectory constraint reinforcement learning module, and the progress deviation verification module.

2. The AI-based labor progress monitoring system according to claim 1, characterized in that, The multi-source labor sensing module includes a fetal heart rate and uterine contraction acquisition unit, a maternal status acquisition unit, a labor examination and recording unit, a bedside ultrasound access unit, and a unified time scale unit. The fetal heart rate and uterine contraction acquisition unit collects the fetal heart rate baseline, fetal heart rate variability, acceleration events, deceleration events, contraction frequency, contraction duration, and contraction intensity. The maternal status acquisition unit collects the mother's heart rate, blood pressure, blood oxygen saturation, and body temperature; The labor examination data entry unit receives information on cervical dilation, cervical effusion, fetal presenting part position, fetal head descent, fetal position, and fetal membrane condition. The bedside ultrasound access unit receives the fetal head progression angle, fetal head-perineal distance, and ultrasound fetal orientation results, and generates data inconsistency markers based on the differences between the bedside ultrasound results and the results of concurrent labor examinations. The unified timescale unit maps continuous monitoring data, labor examination data, bedside ultrasound results, and labor intervention status data to the unified labor time axis based on the server clock.

3. The AI-based labor progress monitoring system according to claim 2, characterized in that, The labor process timeline status construction module generates a time decay reliability status based on the basic reliability of the data source and the time interval from the most recent valid acquisition. It then corrects the time decay reliability status based on data missing status and data mutation status, and generates data reliability coefficients corresponding to various types of monitoring data. The labor process time-series status construction module generates a reliable corrected monitoring result by fusing the actual collection results and short-term trend estimation results with the data reliability coefficient, and generates a reliable fusion vector of the labor process status based on the reliable corrected monitoring result, the data reliability coefficient, the basic data of the parturient, and the labor intervention status data.

4. The AI-based labor progress monitoring system according to claim 3, characterized in that, The labor stage identification module includes a causal time sequence encoder and a stage sequence constraint unit; The causal time-series encoder generates the latent stage probability, active stage probability, second stage probability, and stage transition probability based solely on the current time and the reliable fusion vector of the labor state prior to the current time. The stage sequence constraint unit restricts reverse jumps in labor stages based on the effective labor check results, abnormal data quality status, and previous labor stage identification results, and generates the labor stage identification results and the labor progress prediction trajectory based on the restricted stage probabilities.

5. The AI-based labor progress monitoring system according to claim 4, characterized in that, The trajectory constraint reinforcement learning module includes a first progress benefit evaluator, a second progress benefit evaluator, a safety risk evaluator, a phase consistency reward unit, a safety action masking unit, and a delayed outcome write-back unit. The first progress benefit evaluator and the second progress benefit evaluator generate progress benefit evaluation results based on the same labor status and candidate monitoring actions, and determine the lower result of the two progress benefit evaluation results as the target progress benefit of the candidate monitoring action; The security risk evaluator generates security risk evaluation results based on the missed reporting risk and delayed review risk corresponding to the candidate monitoring actions. The stage consistency report unit generates monitoring action reports based on the accuracy of labor progress prediction, deviation from early identification status, error prompt status, additional inspection burden, deviation from the order of labor stages, and missed reporting status. The safety action shielding unit generates a safety action shielding matrix based on the fetal heart rate status, maternal vital signs status, deviation status of labor progress, and data uncertainty. The trajectory constraint reinforcement learning module generates the target monitoring action from candidate monitoring actions based on the target progress benefit, the safety risk assessment result, the monitoring action reward, and the safety action masking matrix.

6. The AI-based labor progress monitoring system according to claim 5, characterized in that, The candidate monitoring actions include maintaining basic monitoring density, shortening the interval between labor progress reviews, increasing bedside ultrasound review tasks, generating medical and nursing review prompts, and generating high-level medical and nursing review prompts. When the fetal heart rate status, maternal vital signs status, deviation from the labor progress status, and data uncertainty meet the safety shielding conditions, the safety action shielding unit removes the action of maintaining basic monitoring density from the candidate monitoring actions. When medical staff manually increase the monitoring level, the safety action shielding unit prohibits the trajectory constraint reinforcement learning module from generating target monitoring actions that reduce the monitoring level within a set time.

7. The AI-based labor progress monitoring system according to claim 6, characterized in that, After subsequent cervical examination, fetal head descent examination, and bedside ultrasound examination, the delayed outcome write-back unit writes the actual examination results back to multiple labor stages within the risk retrospective window. Based on the consistency between the actual examination results and the corresponding predicted labor progress trajectory, it updates the monitoring action reports corresponding to each labor stage and uses the updated monitoring action reports to adjust the first progress benefit evaluator and the second progress benefit evaluator.

8. The AI-based labor progress monitoring system according to claim 7, characterized in that, The progress deviation verification module generates a labor progress deviation value based on the cervical dilation prediction deviation, the fetal head descent prediction deviation, the uterine contraction response deviation, the confidence level of labor stage identification, and the data uncertainty caused by missing data, expired data, and inconsistent results from multiple sources. It also generates a stable progress status, a pending review status, and a key review status based on the changes in the labor progress deviation value within multiple consecutive monitoring windows. The monitoring feedback output module displays actual examination data, trend estimation data, labor progress prediction interval, main sources of deviation, and suggested review time, and saves the model version, target monitoring actions, medical staff feedback, and subsequent actual examination results.

Images