A method for predicting premature rupture of membranes

Abstract

This invention relates to the field of healthcare informatics and discloses a method for predicting premature rupture of membranes. The method includes: activating a differentiated data input interface based on the hierarchical level of the medical institution and automatically acquiring environmental meteorological parameters; matching clinical indicators and environmental meteorological parameters with a preset high-risk combination logic gate; and generating tiered action instructions and natural language explanation text upon triggering. This invention, through the collaboration of a resource-adaptive interface and rigid logic gates, performs risk identification in low-data environments. Simultaneously, it utilizes a dynamic explanation generation mechanism to construct a transparent decision-making closed loop, preserving algorithm efficiency while enhancing doctor-patient trust, effectively solving the applicability dilemma of traditional probabilistic models in primary healthcare.

Description

Technical Field

[0001] This invention relates to a method for predicting premature rupture of membranes, belonging to the field of healthcare informatics technology. Background Technology

[0002] The current underlying design approach has systemic limitations: on the one hand, the model's strong dependence on data completeness makes it difficult to apply in resource-scarce scenarios such as primary healthcare institutions, forcing doctors to bear the risk of decision-making when key indicators are missing; on the other hand, the characteristic of probability output requiring professional interpretation creates a decision-making black box, which not only exacerbates communication barriers between doctors and patients but also fails to provide clear action guidance.

[0003] A typical dilemma can be seen in emergency response scenarios involving sudden environmental exposure: when pregnant women face acute extreme environmental weather events, existing systems are unable to dynamically perceive the drastic changes in environmental weather parameters, and the lack of stress response mechanisms related to gestational age in static models leads to frequent misjudgments of risk. More importantly, the industry's attempts to optimize accuracy by increasing sensor density or improving algorithm complexity have further exacerbated the contradiction between resource requirements and implementation costs, and the inherent flaw of untraceable decision-making processes has not been eradicated.

[0004] Specifically, existing technologies suffer from three fundamental bottlenecks: 1. A severe mismatch between the unified input interface and grassroots data acquisition capabilities; 2. The temporal evolution characteristics of environmental meteorological parameters and individual physiological states are not incorporated into the risk assessment framework; 3. Probabilistic outputs are difficult to translate into tiered action instructions. Therefore, how to achieve accurate, interpretable, and dynamically adaptive early warning of premature rupture of membranes risk in an environment with highly limited data resources, while simultaneously outputting directly executable clinical action guidelines, becomes the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides a method for predicting premature rupture of membranes. Its main purpose is to solve the problem that it is difficult to achieve accurate, interpretable and dynamically adaptive risk warning in environments with limited data resources, and that its probabilistic output cannot be translated into clear clinical action guidelines.

[0006] To achieve the above objectives, the present invention provides a method for predicting premature rupture of membranes, comprising the following steps: Step a, performing resource level self-identification, and activating differentiated data input interface based on the resource level self-identification result; wherein, for the interface of primary healthcare institutions, the environmental meteorological parameters of the institution's location are automatically obtained in the background through a web application interface.

[0007] Step b involves matching the clinical indicators and environmental meteorological parameters obtained from the data input interface with pre-set high-risk combination logic gates; each logic gate is limited to triggering an alarm when at least two different risk factors simultaneously meet a preset binary condition.

[0008] Step c: When any high-risk combinational logic gate is triggered, a clear hierarchical action instruction is generated and output, and a large language model is invoked to generate a natural language text that can be directly explained to the pregnant woman, using the risk factor that triggered the logic gate as input information.

[0009] Step d: Monitor the rate of change of PM2.5 concentration within a three-hour time window. When the rate of change exceeds the acute shock threshold of 20 micrograms per cubic meter per hour, it is determined to be an acute exposure mode. Under the acute exposure mode, calculate the physiological stress coefficient based on the pregnant woman's current gestational age. The risk level of the graded action instructions is dynamically increased based on the physiological stress coefficient, whereby the physiological stress coefficient is... Furthermore, the physiological stress coefficient is calculated only when the current gestational age is greater than the preset starting gestational age of the physiologically sensitive period. The linear influence coefficient is... This is a risk acceleration factor used to characterize the non-linear growth of risk with gestational age; its value is greater than 1. and All coefficients were pre-calibrated using nonlinear regression analysis of historical clinical data to quantify the impact of gestational age progression on susceptibility to environmental shocks.

[0010] Preferably, in step a, activating the differentiated data input interface includes: if the resource level self-identification result indicates a township health center, the interface will only display three to five basic mandatory input clinical indicator items; if the resource level self-identification result indicates a tertiary hospital, the interface can be switched to expert mode, opening all clinical indicator input interfaces, which contain at least thirty clinical indicators.

[0011] Preferably, in step a, the environmental meteorological parameters automatically acquired in the background by the interface for primary healthcare institutions include PM2.5 concentration and relative humidity.

[0012] Preferably, a high-risk combinational logic gate in step b is defined as follows: the logic gate is triggered when the obtained white blood cell count exceeds a predetermined threshold of 15 x 10^9 per liter and the gestational age exceeds a preset critical gestational age time node.

[0013] Preferably, the graded action instructions generated in step c include: for low-risk levels, the instruction is to follow the routine prenatal examination process; for medium-risk levels, the instruction is to initiate individualized non-clinical interventions, including: activating air purification equipment, supplementing with vitamin C, and avoiding heavy physical labor; for high-risk levels, the instruction is to issue a high-risk alarm, initiate a referral procedure to a higher-level hospital, and implement basic infection prevention measures before referral.

[0014] Preferably, in step c, the risk factor that triggers the logic gate, along with a prompt template used to guide the large language model in generating explanatory text, is sent to the large language model.

[0015] Preferably, when obtaining environmental data via the network application programming interface fails or the returned data timestamp exceeds two hours, the indirect environmental perception assistance method is automatically activated. The indirect environmental perception assistance method includes the following steps: obtaining the average cellular network signal attenuation of the target area through the regionalized, anonymous network quality application programming interface provided by the operator; converting the average cellular network signal attenuation into an environmental risk level through a preset mapping model, the mapping model mapping the signal attenuation value to one of three levels: excellent, moderate, and severe; using the environmental risk level as a substitute environmental meteorological parameter, inputting it into a high-risk combinational logic gate for matching; when the data trigger logic gate of the indirect environmental perception assistance method is used, the system reduces the confidence level of the alarm by a predefined level, or limits the alarm to be confirmed only when at least one clinical positive indicator appears simultaneously.

[0016] Preferably, the method also includes a dual-channel decision arbitration method based on the principle of symptom priority. The method includes: receiving a preliminary risk level suggestion from the main warning module in step c; obtaining information input by medical staff regarding whether the pregnant woman currently has typical clinical manifestations, including vaginal fluid discharge, posterior fornix effusion, and positive amniotic fluid crystals; generating and outputting a final action instruction based on preset arbitration rules, combined with the preliminary risk level suggestion and key clinical manifestation information; wherein, if the pregnant woman has vaginal fluid discharge combined with posterior fornix effusion and positive amniotic fluid crystals, then regardless of the preliminary risk level suggestion, the final action instruction will forcibly output the highest level alarm, extremely high risk, suspected rupture of membranes, please immediately follow the premature rupture of membranes procedure.

[0017] Compared with existing technologies, the beneficial effects of this invention are: 1. The self-identification mechanism at the medical institution level automatically adapts to the complexity of data input. In grassroots scenarios, only extremely simple clinical indicators need to be input, while environmental meteorological parameters are silently acquired in the background. This design transforms the traditional massive data requirements into risk factor capture, enabling the binary triggering mechanism of high-risk combination logic gates to operate stably in low-resource environments. Grassroots doctors do not need to interpret complex probability models to obtain clear action instructions, so that action instructions can still be generated when data indicators are limited.

[0018] 2. High-risk combinational logic gates replace probability-weighted calculations with rigid conditions that simultaneously satisfy multiple factors, directly outputting binary action commands. Simultaneously, a large-scale language model is invoked to convert triggering factors into natural language explanations, forming a command-explanation closed loop. This mechanism makes the system's output action commands and their triggering factors visible to the user, preserving the machine's risk identification capabilities while supporting doctor-patient communication through interpretable output, avoiding trust barriers in algorithmic decision-making. Furthermore, by continuously monitoring the rate of change in environmental meteorological parameters to identify acute exposure events and dynamically calculating physiological stress coefficients based on gestational age, this mechanism elevates static threshold judgment to spatiotemporal dynamic response, enabling the system to distinguish the risk differences between chronic pollution and sudden environmental shocks. Combined with the gestational age coefficient for flexible adjustment of alarm levels, the risk level is adjusted according to gestational age while maintaining the simplicity of the basic scheme.

[0019] 3. The independently operating clinical symptom arbitration channel has the right to veto the main logic channel, ensuring that key signs such as premature rupture of membranes have the highest processing priority. When the environmental data source fails, the cellular signal attenuation mapping model provides alternative risk assessment and avoids false alarms through the confidence downgrading mechanism. The two constitute a redundant verification system, enabling the system to output instructions according to the preset symptom priority rules or confidence downgrading rules even in the event of data loss or noise interference. Attached Figure Description

[0020] Figure 1 This is a decision logic architecture diagram of the risk warning system of the present invention.

[0021] Figure 2 This is a system state transition path diagram for three typical scenarios of the present invention.

[0022] Figure 3 This is a diagram illustrating the system decision-making logic and risk state transition path of the present invention. Detailed Implementation

[0023] To make the purpose and advantages of this technical solution clearer, a method for predicting premature rupture of membranes is described in detail. Its overall operation process begins with a resource-level self-identification step that does not require manual intervention, and then goes through core stages such as high-risk combination logic gate matching, generation of graded action instructions and interpretation texts, and dynamic risk adjustment based on environmental meteorological parameters, ultimately forming a closed-loop and adaptive clinical decision support path.

[0024] When healthcare workers activate this method on their terminal devices, the system first automatically identifies the resource level by reading pre-set institution authentication information or parsing specific identifiers of the current network environment. Given the objective differences in data collection capabilities and needs among different levels of medical institutions, the system activates differentiated data input interfaces based on the identification results. If the institution is determined to be a township health center with relatively limited data resources, the interface will present a minimalist interface focusing on the input of three to five basic mandatory clinical indicators, such as gestational age, body temperature, and white blood cell count. Simultaneously, the system silently acquires real-time environmental meteorological parameters corresponding to the institution's geographical location via a web application interface in the background, particularly PM2.5 concentration and relative humidity, thus aggregating multi-source data without increasing the operational burden on primary healthcare workers. Conversely, if the identification result is a well-resourced tertiary hospital, the interface provides an option to switch to expert mode, which opens all input interfaces containing at least thirty clinical indicators. This adaptive mechanism effectively solves the inherent problem of insufficient applicability of a unified data model in different resource environments. Subsequently, the system sends the clinical indicators obtained from the interface and the environmental meteorological parameters obtained from the backend into a core risk discrimination module. The cornerstone of this module is a series of high-risk combinational logic gates based on deterministic rules. In this prediction model, various risk factors are not equivalent. Based on clinical experience, physiological and clinical factors are the core basis for predicting premature rupture of membranes, contributing approximately 85% to the model results; while environmental meteorological factors are considered secondary or precipitating factors, contributing approximately 15%. Unlike traditional probability models that use weighted summation compensatory logic, each logic gate here... All adopt a non-compensatory design, which requires that at least two different risk factors must simultaneously meet their preset binary conditions to be triggered. This design directly avoids false positives caused by abnormalities in a single indicator. For example, the triggering condition of one of the logic gates is strictly limited to: when the obtained white blood cell count exceeds the prescribed threshold of 15 x 10^9 per liter as set according to clinical guidelines, and the pregnant woman's gestational age also exceeds the preset key time node of gestational age, the logic gate is triggered. This series of rigid logic based on solidified medical knowledge is used to generate action instructions.

[0025] To enable those skilled in the art to implement this method, the construction method of high-risk combinational logic gates and the generation procedure of explanatory text are further explained here. For example, one logic gate can be limited to being triggered when the relative humidity of the air obtained from the background is greater than 70% for six consecutive hours and the pregnant woman has a history of recurrent vaginitis; another logic gate can be limited to being triggered when the C-reactive protein level exceeds 10 mg / L and the body temperature is higher than 37.5 degrees Celsius. When either logic gate is triggered, the system sends the triggering factor data and a structured prompt template to the large language model interface. This template includes a role definition field for instructing the model to communicate as a professional, a risk factor list field for loading the specific factors that are being triggered, and a constraint rule field that restricts the model output to not contain diagnostic conclusions and not recommend specific treatment plans through text instructions. If the interface fails to return valid data within five seconds, the system abandons the call and retrieves a standardized explanatory statement from the local database that uniquely corresponds to the code of the currently triggered logic gate and displays it directly. Among them, the physiological stress coefficient used for dynamic risk adjustment is... The calibration procedure for the parameter K in this study begins by selecting a localized historical dataset containing at least one thousand confirmed cases. Outliers are then removed from the relative hazard ratio data points for each case using the interquartile range method. Subsequently, a nonlinear regression analysis method (e.g., nonlinear least squares) is used to fit the preprocessed hazard ratios to the gestational age data to find the optimal combination of parameters that best describes the nonlinear relationship curve. The goal of this analysis is to simultaneously calibrate the linear influence coefficient. With risk acceleration factor Finally, the goodness of fit of this nonlinear model is calculated (e.g., adjusted). The model's accuracy is evaluated using a goodness-of-fit value. Only when this goodness-of-fit exceeds a preset threshold (e.g., 0.5) is the calibration result considered accurate. and It was only after this process that it was finally adopted; in addition, the segmentation threshold in the indirect environment perception mapping model and The model was obtained by applying ordered logistic regression analysis to historical signal attenuation values ​​and PM2.5 concentration levels marked according to national air quality standards. The classification accuracy of the model was internally validated using the ten-fold cross-validation method, and the set of thresholds with the best F1 score on the validation set was used as the final deployment parameters.

[0026] In the risk assessment module of this technical solution, the specific construction of a series of high-risk combination logic gates follows the principle of joint triggering by multiple risk factors. Among all risk factors, infection-related indicators (such as white blood cell count, C-reactive protein, and history of recurrent vaginitis) should be given priority, as intrauterine infection is the most significant core cause of premature rupture of membranes. Specifically, among all risk factors, infection-related physiological and clinical indicators should be given priority, because according to clinical experience, physiological factors are the core basis for predicting premature rupture of membranes, contributing approximately 85% to the model results; while environmental and meteorological factors are considered secondary or triggering factors, contributing approximately 15%. For example, in addition to the aforementioned examples, a logic gate can... The system is triggered when the relative humidity of the air acquired by the backend exceeds a high humidity threshold of 70% for six consecutive hours and the pregnant woman has a history of recurrent vaginitis. This setting aims to identify situations where persistently high humidity creates conditions for pathogen growth, thereby increasing the risk of opportunistic infections. To ensure the regional adaptability and accuracy of this risk identification, the high humidity threshold is not a preset fixed value. The 70% threshold is not a fixed value, but rather is determined by performing receiver operating characteristic (ROC) curve analysis on at least two thousand anonymous historical obstetrics and gynecology visit records and concurrent meteorological records obtained before deployment in the target area. This ROC curve analysis is used to find the optimal critical point of a binary classification model by calculating the true positive rate and false positive rate. The calculation method identifies the humidity cutoff point with the highest correlation to the seasonal inflection point of vaginitis consultation rate; inflammatory markers can be used together or selectively with indicators such as white blood cell count (WBC) for inflammatory risk assessment. The CRP threshold here is also determined by the procedure, which involves screening all confirmed cases of chorioamnionitis in the local database, tracing back their CRP values ​​within 48 hours before membrane rupture, constructing their probability density function, and taking the 75th percentile value of this function distribution as the warning threshold; once any high-risk combinational logic gate is triggered, the system immediately enters the instruction generation and interpretation stage; the system first accurately matches and outputs the corresponding graded action instruction from a preset instruction library based on the specific logic gate code that was triggered. These instructions are clearly categorized according to risk levels, covering low-risk cases (simply follow routine prenatal checkups), potential risk warnings (enhanced observation, retesting key indicators after 12 hours or consulting a senior physician), and even high-risk alerts (recommending immediate referral to a higher-level hospital). Simultaneously with issuing the instructions, the system triggers the specific risk factor of the corresponding logic gate, along with a prompt template for generating standardized explanatory text, and sends them to a large language model. Based on the received precise input, the model instantly generates a natural language text that can be directly communicated to the pregnant woman, objectively stating the current risk combination. This integrates risk identification, instruction delivery, and transparent explanation, providing direct and complete support for clinical communication.

[0027] Furthermore, to address the acute impact of sudden environmental events on the physiological state of pregnant women, this method integrates a risk regulation mechanism that dynamically senses environmental meteorological parameters and adapts to physiological stages. The system continuously monitors environmental meteorological parameters in the background and focuses on calculating their rate of change within a three-hour time window. When this rate of change exceeds an acute impact threshold, such as 20 micrograms per cubic meter per hour, calibrated based on large-scale epidemiological data, the system determines that it has entered an acute exposure mode. In this mode, if the pregnant woman's current gestational age is within a physiologically sensitive period that is greater than the preset starting gestational age, the system initiates dynamic risk regulation based on the physiological stress formula. Calculating the physiological stress coefficient , where constant It is a pre-calibrated coefficient, derived through regression analysis of historical clinical data, used to quantify the impact of gestational age progression on susceptibility to environmental shocks, and is relevant to the physiological stress coefficient. Parameters in The calibration procedure for this value begins with data preprocessing, specifically using the interquartile range (IQR) method on the relative hazard ratio data points of the enrolled cases. IQR is a statistical technique that uses the difference between the first and third quartiles of the dataset to identify outliers. and The first and third quartiles, respectively, are used to construct... Data points within a given interval are considered outliers and removed. Subsequently, a nonlinear regression analysis method (e.g., nonlinear least squares) is used to fit the preprocessed hazard ratios to the gestational age data to find the optimal combination of parameters that best describes the nonlinear relationship curve. The goal of this analysis is to simultaneously determine the linear influence coefficient. With risk acceleration factor Finally, the goodness of fit of this nonlinear model is calculated (e.g., adjusted). The model's accuracy is evaluated using a goodness-of-fit value. Only when this goodness-of-fit exceeds a preset threshold (e.g., 0.5) is the calibration result considered accurate. and Only then was it ultimately adopted; in addition, to address the technical issue that continuous changes in the external environment and data distribution may lead to a decrease in model effectiveness, the system has a built-in periodic self-calibration mechanism: whenever the system accumulates 500 new cases of premature rupture of membranes diagnosed by the gold standard within the authorized medical network or the system's continuous running time reaches twelve months, it will automatically re-execute the parameter calibration as a background task. And threshold for indirect environment perception mapping model , A complete calibration and analysis procedure is established to enable continuous adaptation to changes in the external environment and data distribution.

[0028] Example 1: In a specific application scenario, a township health center's obstetrics clinic faced a clinical decision-making challenge one afternoon. A pregnant woman at 35 weeks gestation presented with symptoms of discomfort. Her clinical signs were nonspecific. Gynecological examination revealed no vaginal discharge, routine examination of secretions showed no abnormalities, and the amniotic fluid crystallization test was negative, initially ruling out a ruptured membranes diagnosis. However, her blood test showed a slightly elevated white blood cell count, but it did not reach the usual infection warning threshold. At this point, the attending physician faced the challenge of accurately diagnosing the condition: unable to determine whether the pregnant woman's nonspecific symptoms were caused by a subclinical infection sufficient to cause premature rupture of membranes; simultaneously, the air quality in the area was poor due to a nearby forest fire. The situation deteriorated rapidly within a short period, with a non-linear surge in atmospheric particulate matter concentration. When the doctor entered the pregnant woman's limited basic clinical indicators into the prediction system of this technical solution, the system, which had already completed resource-level self-identification, immediately presented a highly simplified data input interface containing only the core required fields. At the same time, the system's backend automatically obtained PM2.5 concentration and relative humidity data from the local environmental monitoring station via a web application interface. In the initial logic matching, because the pregnant woman's clinical indicators failed to independently trigger any high-risk combination logic gates, the system only output a potential risk warning: a preliminary instruction to strengthen observation. This situation reflects the inherent limitations of risk assessment in data-sparse scenarios.

[0029] However, the system's environmental meteorological parameter dynamic sensing module operated in parallel during this period. By calculating the rate of change of PM2.5 concentration data within a three-hour time window, it identified that the growth rate of this parameter had exceeded the acute shock threshold of 20 micrograms per cubic meter per hour, and the system immediately determined to enter the acute exposure mode. Given that the pregnant woman's current gestational age of 35 weeks was greater than the preset physiological sensitive period starting week boundary, the system automatically activated the physiological stress coefficient formula. The calculation yields the coefficient values. Used as a dynamic risk gain factor, this coefficient provides quantitative physiological evidence for reassessing the initial risk level. It logically couples the potential risk, originally composed of atypical clinical indicators, with the external stressor of acute environmental shock. This causes the overall risk level to exceed a higher internal alarm threshold, prompting the system to dynamically escalate the original action instruction to a high-risk alarm: recommending immediate referral to a higher-level hospital. Simultaneously, a large-scale language model is invoked to generate a clear natural language explanation text for all factors triggering this alarm escalation—namely, elevated white blood cell count and acute environmental exposure—to aid in doctor-patient communication. The final clinical intervention decision... The strategy is not based on any single, significant positive sign, but rather on a logical connection between non-specific clinical data and an independently monitored acute environmental stress event. This connection is revealed and quantified through a system framework that integrates static data input and dynamic risk adjustment. Thus, in a primary healthcare environment with incomplete information, it facilitates timely intervention through risk assessment. Its significance lies not only in providing pregnant women with guidance on self-prevention and seizing the right time for precise medical treatment, but also in providing key technical support for optimizing referral pathways and achieving a rational allocation of scarce medical resources for the entire primary healthcare system.

[0030] Example 2: This example verifies the effectiveness of the dynamic sensing of environmental meteorological parameters and the adaptive adjustment mechanism of physiological stages in this technical solution through a controlled simulation experiment, particularly examining the response differences between chronic and acute environmental exposure modes. For this verification, the experimental platform was constructed as a pure software simulation environment, consisting of a virtual pregnant woman data generator, a programmable environmental condition simulator, and a system instance fully integrated with this technical solution to be tested. The virtual pregnant woman data generator was set to continuously output a set of standardized clinical data from pregnant women at 32 weeks of gestation. Key indicators such as white blood cell count were set at a subclinical level slightly above normal but below the trigger threshold of any single high-risk combinational logic gate, thereby establishing a stable and non-built-in clinical... The baseline state of the alarm; the total simulation duration is set to six hours. This period is designed to ensure that after an environmental event is introduced in the middle of the simulation, the system has a complete three-hour time window to execute its rate of change monitoring and judgment logic; the experiment is conducted in parallel with two groups. The environmental condition simulator of control group A maintains a high but constant PM2.5 concentration of 85 micrograms per cubic meter throughout the six-hour simulation period, with a concentration change rate of zero, to simulate a chronic exposure scenario; the experimental group B maintains a PM2.5 concentration of 20 micrograms per cubic meter for the first three hours. At the exact three-hour mark, the concentration linearly rises to 90 micrograms per cubic meter within one hour and remains at that level thereafter. The concentration change rate within a specific time period exceeds 20 micrograms per cubic meter per hour, which is the acute impact threshold, to simulate an acute exposure scenario.

[0031] After the experiment started, the system received data generated by two sets of simulators and executed its internal logic. During the entire operation of control group A, although the PM2.5 concentration remained high, its rate of change was always zero, so the system's environmental meteorological parameter dynamic sensing module did not determine that it had entered acute exposure mode, and the physiological stress coefficient... The system output action instructions were not calculated and remained at the lowest level based on subclinical indicators, namely, potential risk warning: strengthen observation; In experimental group B, after the simulation time reached four hours, the system's monitoring module detected that the rate of change of PM2.5 concentration had exceeded the threshold. The system determined that it had entered the acute exposure mode and triggered the subsequent risk regulation mechanism. Its key state evolution data are shown in Table 1.

[0032] Table 1: Comparison of system status under different exposure modes.

[0033]

[0034] After experimental group B entered acute exposure mode, the system simultaneously processed data from two sets of virtual pregnant women (32 weeks and 37 weeks) generated in parallel. In the parameter calibration of this embodiment, the preset starting gestational age of the physiologically sensitive period was set to 28 weeks. The system then calculated and performed a judgment based on the fact that the current gestational age was greater than the preset starting gestational age of the physiologically sensitive period. This logic ensured the coefficient... The calculation uses the following segmented method: if the current gestational age is less than the initial gestational age, then... The value remains at 1; if the current gestational age is greater than the initial gestational age, the system activates the risk adjustment mechanism; the system applies this segmentation logic to the two sets of parallel data in this embodiment, and the specific process is as follows: For the first set (32 weeks of gestation): the system determines that 32 weeks > 28 weeks, triggers risk adjustment, and calls the formula. To perform a calculation, substitute the numerical values; that is, perform the calculation. For the second group (37 weeks of pregnancy): the system determines that 37 weeks is greater than 28 weeks, triggering risk adjustment, and calls the same formula to calculate... The starting gestational week (28 weeks) here is a statistically significant inflection point. A comparison shows that... The calculation base ( (far greater than) The calculation base ( Finally, the physiological stress coefficient was calculated. ( and (All values ​​greater than 1) were used as a weighting factor, increasing the baseline risk values ​​of the two groups composed of subclinical indicators. This caused the comprehensive risk assessment results of the two groups to cross the preset alarm threshold, and the final action instruction was dynamically upgraded to a high-risk alarm: it is recommended to immediately contact a higher-level hospital for referral. This difference in results directly links the upgrade of the system decision to the capture of the temporal evolution characteristics of environmental meteorological parameters, rather than their static absolute values. The data from this experiment show that the dynamic adjustment mechanism based on the rate of change trigger in the prediction method can distinguish between two different environmental exposure modes when receiving the same subclinical pregnant women data. The mechanism maintains the baseline alarm level when facing static high concentrations of environmental meteorological parameters, while raising the alarm level when encountering dynamic environmental meteorological parameter shocks exceeding the preset threshold. This confirms the ability of the mechanism to operate as designed.

[0035] To further verify the necessity and non-obviousness of the dynamic sensing and physiological stage adaptive adjustment mechanism of environmental meteorological parameters based on the rate of change described in this invention, the following comparative example 1 is established.

[0036] Comparative Example 1: This comparative example is intended to provide a direct comparison with Example 2 to illustrate the non-obviousness of the technical solution of the present invention. The experimental platform, virtual pregnant woman data generator, baseline state (using the same 32-week and 37-week gestational data as in Example 2, with key clinical indicators at subclinical levels), and total simulation time (six hours) used in this comparative example are exactly the same as in Example 2. The essential difference lies in replacing the dynamic sensing of environmental meteorological parameters and the physiological stage adaptive adjustment mechanism in Example 2 with a conventional environmental risk early warning method based on a fixed threshold in the art. Specifically, this conventional method does not calculate the rate of change of PM2.5 concentration, nor does it introduce a physiological stress coefficient. Instead, a fixed, absolute environmental alert threshold is set, defining a PM2.5 concentration exceeding 75 micrograms per cubic meter as a trigger condition for a high level of environmental risk, based on relevant public health guidelines.

[0037] The experimental process was as follows: an experimental group (named Comparative Experimental Group C) was established, maintaining a PM2.5 concentration of 20 micrograms per cubic meter for the first three hours. At the exact three-hour mark, the concentration linearly increased to 70 micrograms per cubic meter within one hour and remained at that level thereafter. This scenario simulated a common acute exposure event in the real world, where the pollution level was significant but had not yet reached the standard for severe pollution. Throughout the entire operation of Comparative Experimental Group C, the clinical data of the virtual pregnant woman received by the system remained at a subclinical level, so its preliminary risk level based on clinical indicators remained a potential risk warning. In the fourth hour of the simulation, although the PM2.5 concentration experienced a rapid increase from 20 micrograms per cubic meter to 70 micrograms per cubic meter, its absolute value of 70 micrograms per cubic meter did not reach the preset fixed alarm threshold of 75 micrograms per cubic meter. Therefore, the environmental risk warning method based on the fixed threshold was not triggered, and the action instructions output by the system remained at the lowest level. Its key state evolution data are shown in Table 2.

[0038] Table 2: System status table using the conventional fixed threshold method.

[0039]

[0040] The experimental results of Comparative Example 1 show that the conventional early warning method in the art, based on a fixed absolute value threshold of environmental parameters, cannot identify the inherent sudden risk in an acute environmental exposure event that does not reach a high-risk absolute value but has a significant rate of change, and therefore fails to effectively correct the initial risk level. In contrast, the results of Example 2 confirm that the present invention, by monitoring the rate of change of environmental parameters and combining it with the dynamic calculation of physiological susceptibility coefficients based on gestational age, can accurately capture such acute shock risks and dynamically raise the alarm level. The comparison of the two demonstrates the significant technical progress and non-obviousness of the present invention's technical solution in addressing the problem that existing technologies cannot effectively cope with sudden environmental exposures.

[0041] Example 3: This example combines Figures 1 to 3 A method for predicting premature rupture of membranes is described, such as... Figure 1As shown, medical staff input clinical indicators, while the environmental monitoring data interface and network quality application interface provide environmental meteorological parameters and network signal attenuation values ​​as backup data sources, respectively. These data are fused into fused data in the data acquisition and interface adaptation module and then enter the risk factor matching module. This module performs logic gate matching based on the high-risk combination logic gate rules stored in database D1. The preliminary matching results flow to the decision arbitration module, which combines the key clinical manifestation information input externally to arbitrate and generate arbitration instructions. On the other hand, real-time PM2.5 data is sent to the dynamic risk adjustment module, which performs calculations based on the calibration parameters in database D2 to generate risk adjustment coefficients. The instruction and interpretation generation module integrates the arbitration instructions and risk adjustment coefficients to generate graded action instructions and feeds them back to the medical staff. At the same time, the triggering factors are sent to the large language model, which generates natural language interpretations based on the original interpretation text and transmits them to the pregnant woman.

[0042] like Figure 2 As shown, with the decision time point as the horizontal axis and the system state as the vertical axis, three characteristic curves intuitively illustrate three typical paths of system state evolution over time. The normal monitoring curve shows that the system enters the monitoring state from the initial state at decision time point 2 and remains in this state thereafter. The alarm trigger curve depicts the evolution process of the system directly entering the potential risk state from the initial state and escalating to the high-risk state at decision time point 4. The emergency veto curve represents the triggering of the dual-channel decision arbitration mechanism based on the symptom priority principle, indicating that the system state will skip the intermediate risk level and be forcibly escalated to the extremely high-risk state.

[0043] like Figure 3 As shown, after receiving clinical and environmental data, the system enters a monitoring state for preliminary analysis. After the preliminary matching is completed, if a high-risk combination logic gate is triggered, the system enters a high-risk alarm state. If no high-risk logic gate is triggered but subclinical indicators are present, the system enters a potential risk state. In this state, if the PM2.5 change rate exceeds the acute impact threshold, the system will enter a dynamic risk level escalation state, enter an acute exposure mode, and calculate the physiological stress coefficient. In addition, in the potential risk or high-risk alarm state, once a key clinical manifestation is detected, the system will forcefully enter an extremely high-risk state, that is, the alarm for suspected rupture of membranes will be forcibly triggered by symptom priority arbitration. Finally, all high-risk, extremely high-risk, and dynamically escalated risk states lead to the endpoint of generating the final instruction.

[0044] Example 4: This example aims to provide a detailed engineering description of the indirect environmental perception assistance method and its core mapping model, along with the calibration process for key risk parameters, in this technical solution. This addresses the technical problem of how the system maintains its environmental risk assessment function when direct environmental monitoring data sources are lacking. In a specific deployment scenario, this technical solution needs to be applied to an area with good cellular network coverage but unable to obtain real-time official environmental monitoring data, or where the response delay of available interfaces exceeds two hours. This situation makes immediate assessment of acute environmental exposure impossible. If the system cannot activate an effective alternative environmental perception solution at this time, its risk assessment capability will rely solely on clinical indicators and will be unable to address risks induced by unperceived environmental factors. To address this... To address this challenge and ensure the system's robustness, an offline mapping model construction and parameter calibration process for the target area must be executed before formal deployment. This process begins with data preparation, the goal of which is to obtain two types of time-synchronized historical datasets covering at least twelve months of the region. One type is the hourly average cellular network signal attenuation obtained through a regionalized, anonymized network quality application programming interface provided by the operator. The other type is the hourly historical record of PM2.5 concentration obtained from the nearest and generally reliable official environmental monitoring station in the region. After acquiring the data, the two sets of data are timestamped and missing values ​​are imputed to form a coupled dataset containing signal attenuation values ​​and their corresponding PM2.5 concentrations at that time.

[0045] The next step is to construct the mapping model. The core of this stage is to map continuous signal attenuation values ​​to three discrete environmental risk levels: excellent, moderate, and severe. First, based on national air quality standards, the PM2.5 concentration values ​​in the coupled dataset are classified and labeled. One specific criterion is to label records with concentrations less than 35 micrograms per cubic meter as excellent, those between 35 and 75 micrograms per cubic meter as moderate, and those greater than 75 micrograms per cubic meter as severe. Then, using the average signal attenuation as the independent variable and the labeled environmental risk level as the dependent variable, ordered logistic regression analysis is applied. The goal of this analysis is to find the statistically optimal segmentation threshold that classifies different signal attenuation values ​​into their corresponding risk levels. The analysis results show that when the average signal attenuation is below a certain threshold… At that time, the probability of the corresponding environmental risk level being excellent exceeds 90%, while when the average signal attenuation is higher than another threshold... At that time, the probability of the corresponding environmental risk level being severe exceeds 90%, and the threshold... and This constitutes the core decision boundary of the pre-defined mapping model.

[0046] At the same time, using this coupled dataset, the physiological stress coefficient was analyzed. Parameters in The acute shock threshold was localized and calibrated. All cases of premature rupture of membranes (PROM) were screened from the dataset, and PM2.5 concentration changes in the 72 hours prior to the event were retrospectively analyzed. The 95th percentile of this distribution was determined as the acute shock threshold for that region. Furthermore, these cases were grouped by gestational age, and the relative risk ratios of PROM in different gestational age groups under the same acute shock were calculated. Then, nonlinear regression analysis was used to fit this risk ratio to the gestational age data to simultaneously determine the linear influence coefficient that best describes this nonlinear relationship. With risk acceleration factor This process provides a traceable statistical basis for setting key parameters. Through the above offline calibration process, a data-driven mapping model with clear decision boundaries is constructed for the indirect environmental perception auxiliary method, and the core parameters in dynamic risk assessment are locally calibrated. This enables the environmental risk assessment module of this technical solution to continue to operate based on a quantifiable and verifiable alternative data source when deployed in areas with limited data.

[0047] Example 5: In a specific system deployment and initialization application scenario, to ensure the functional integrity and operational safety of this technical solution before its clinical use, a standardized pre-test and model verification procedure needs to be executed. This procedure runs automatically when the system is first started in a medical institution or after a major software update. It first performs an external interface connectivity self-test. The system will sequentially attempt to call the external network application interfaces it depends on, including the main environmental meteorological parameter acquisition interface and the large language model service interface. It will verify the validity of the network link and service authorization by verifying the authentication status code and test data packet returned. If any interface connection fails or times out, a clear non-blocking mark will be displayed on the system's main interface, indicating the specific restricted functional module.

[0048] After the interface self-check passes, the procedure further solidifies and verifies the interaction logic of the large language model. When the system calls the large language model to generate explanatory text, it sends a structured data object to the interface. This object contains a field that defines the model's role, a data field that contains only the list of risk factors involved in this trigger, and a control field that contains multiple mandatory constraint rules. These constraint rules instruct the model not to provide diagnostic conclusions, not to recommend treatment plans, and to make objective statements based solely on the provided data factors. To deal with abnormal situations such as the large language model interface call failure or non-compliant returned content, the system has a built-in deterministic text replacement mechanism. If the interface call fails to return a valid result within a preset five seconds, the system will automatically abandon the call and display a standardized explanatory statement that corresponds one-to-one with the triggered high-risk combinational logic gates and is pre-stored locally at the original location of the explanatory text in the user interface.

[0049] The procedure also includes the quantitative setting of specific operating parameters. For the confidence adjustment of alarms triggered by indirect environmental perception assistance methods, the system internally defines the confidence level as an integer level from 1 to 5. The initial confidence level of the standard alarm triggered by the direct data source is 5. When the system enables the indirect environmental perception assistance method due to the failure to acquire the main environmental data, it will subtract 1 from the initial value of the alarm and adjust it to 4 when the alarm is finally generated, thereby quantifying the change in confidence. By executing this whole set of pre-procedures, the system's functional dependencies and behavior patterns under abnormal conditions are pre-verified and defined before entering actual operation, so as to ensure its operational reliability in complex clinical environments.

[0050] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method of predicting preterm premature rupture of membranes, characterized by, The method comprises the following steps: Step a, perform resource level self-identification, and activate the differentiated data input interface according to the resource level self-identification result; wherein, for the interface of primary medical institutions, the environmental meteorological parameters of the location of the institution are automatically obtained through the network application program interface in the background; step b, match the clinical indicators and environmental meteorological parameters obtained from the data input interface with the pre-set high-risk combination logic gate; each logic gate limits at least two different sources of risk factors to trigger an alarm when the pre-set binary condition is met; step c, when any high-risk combination logic gate is triggered, generate and output a clear graded action instruction, and call a large language model, input the risk factors of the triggered logic gate as input information, and generate a natural language text that can be directly explained to pregnant women; step d, monitor the change rate of particulate matter PM2.5 concentration in a continuous three-hour time window, and when the change rate exceeds the acute impact threshold of twenty micrograms per cubic meter per hour, it is judged as an acute exposure mode; in the acute exposure mode, calculate the physiological stress coefficient according to the current gestational age of the pregnant woman , and dynamically increase the risk level of the graded action instruction according to the physiological stress coefficient, wherein the physiological stress coefficient , and only when the current gestational age is greater than the pre-set physiological sensitive period starting gestational age, the physiological stress coefficient is calculated, wherein is a linear influence coefficient, is a risk acceleration coefficient for representing the nonlinear growth characteristics of risk with gestational age, and the value is greater than 1, and are pre-calibrated by nonlinear regression analysis of historical clinical data, which are used to quantify the influence of gestational age growth on environmental impact susceptibility together.

2. The method of predicting preterm premature rupture of membranes according to claim 1, wherein, In step a, the activation of the differentiated data input interface comprises: if the resource level self-identification result indicates a township hospital, the interface only displays three to five basic mandatory input clinical indicator input items; if the resource level self-identification result indicates a three-A hospital, the interface can be switched to an expert mode, and all clinical indicator input interfaces are opened, which contain at least thirty clinical indicators.

3. The method of predicting preterm premature rupture of membranes according to claim 1, wherein, In step a, the environmental meteorological parameters obtained automatically in the background by the interface for primary medical institutions include the concentration of particulate matter PM2.5 and the relative humidity of air.

4. The method of predicting preterm premature rupture of membranes according to claim 1, wherein, The high-risk combination logic gate in step b is defined as: the logic gate is triggered when the obtained white blood cell count value exceeds the specified threshold value of 15 multiplied by ten to the ninth power per liter and the gestational age exceeds a preset gestational age key time node.

5. The method of predicting preterm premature rupture of membranes according to claim 1, wherein, The graded action instructions generated in step c include: for the low-risk level instruction, following the routine antenatal examination process; for the medium-risk level instruction, starting individualized non-clinical intervention, which includes: enabling air purification equipment, supplementing vitamin C, and avoiding heavy physical labor; for the high-risk level instruction, outputting a high-risk alarm, starting the referral procedure to a higher-level hospital, and performing basic infection prevention measures before referral.

6. The method of predicting preterm premature rupture of membranes according to claim 1, wherein, In step c, the risk factors that trigger the logic gate are sent to the large language model together with a prompt template for guiding the large language model to generate explanatory text.

7. The method of predicting preterm premature rupture of membranes according to claim 1, wherein, When the environmental data fails to be obtained through the network application program interface or the data timestamp returned exceeds two hours, an indirect environmental perception auxiliary method is automatically started, which comprises the following steps: obtaining the average value of cellular network signal attenuation in the target area through the regional and anonymous network quality application program interface provided by the operator; converting the average value of cellular network signal attenuation into an environmental risk level through a preset mapping model, which maps the signal attenuation value into one of the three levels of excellent, medium, and poor; inputting the environmental risk level as a substitute for the environmental meteorological parameter into the high-risk combination logic gate for matching; when the data obtained by using the indirect environmental perception auxiliary method triggers the logic gate, the system reduces the confidence level of the alarm by a predefined level, or limits that at least one clinical positive indicator must appear simultaneously to confirm the alarm.

8. The method of predicting preterm premature rupture of membranes according to claim 1, wherein, It also comprises a two-channel decision arbitration method based on the symptom priority principle, which comprises: receiving a preliminary risk level suggestion from the main warning module of step c; obtaining the input by medical personnel about whether the pregnant woman currently has typical clinical manifestations, including liquid flowing out of the vagina, posterior fornix fluid, and amniotic fluid crystallization positive; generating and outputting a final action instruction according to the preset arbitration rules combined with the preliminary risk level suggestion and the key clinical manifestation information; wherein, if the pregnant woman has liquid flowing out of the vagina combined with posterior fornix fluid and amniotic fluid crystallization positive, regardless of the preliminary risk level suggestion, the final action instruction is forced to output the highest level of alarm, extremely high risk, suspected fetal membrane rupture, please handle immediately according to the fetal membrane rupture process.

Images