Method and apparatus for generating a monitoring report

Abstract

The application relates to the technical field of data processing, in particular to a monitoring report generation method and device. The method comprises the following steps: inputting a monitoring data sequence into a hybrid neural network model to extract local morphological features, and then identifying key physiological features to obtain a feature analysis result; generating a primary report according to the feature analysis result and acquired comprehensive information of a monitoring object; matching each piece of information in the primary report with a preset expert experience rule library to generate at least one risk prompt information; constructing structured context information according to the primary report and the at least one risk prompt information; and inputting the structured context information into a retrieval enhanced large language model to generate a monitoring report conclusion in a natural language form. The monitoring report conclusion generated by the application is more accurate.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and device for generating monitoring reports. Background Technology

[0002] In existing technologies, automated monitoring is achieved based on physiological signals, using neural networks to extract physiological features. Monitoring reports are then generated based on these features. However, these reports suffer from low accuracy and readability; in particular, inaccurate interpretations of fetal heart rate monitoring reports can have significant negative impacts. Summary of the Invention

[0003] In view of this, embodiments of this application provide a monitoring report generation method and device, which can effectively solve the problem of low accuracy in generating monitoring reports in the prior art.

[0004] In a first aspect, embodiments of this application provide a method for generating a monitoring report, including:

[0005] The monitoring data sequence is input into a hybrid neural network model to extract local morphological features, thereby identifying key physiological features and obtaining feature analysis results;

[0006] Based on the feature analysis results and the comprehensive information of the monitored objects obtained, a preliminary report is generated;

[0007] Each piece of information in the primary report is matched with a preset expert experience rule base to generate at least one risk warning message;

[0008] Construct structured contextual information based on the primary report and at least one of the risk warning messages;

[0009] The structured context information is input into the retrieval-enhanced large language model to generate a monitoring report conclusion in natural language form.

[0010] In some embodiments, the monitoring data sequence includes a first type of data sequence and a second type of data sequence;

[0011] The process of inputting the monitoring data sequence into a hybrid neural network model to extract local morphological features, thereby identifying key physiological features and obtaining feature analysis results includes:

[0012] The first type of data sequence and the second type of data sequence are input into the first neural network to identify the local morphological features;

[0013] The local morphological features are input into a second neural network to track changes in the local morphological features, identify at least one of the key physiological features, and generate the feature analysis results based on at least one of the key physiological features.

[0014] In some embodiments, generating the feature analysis result based on at least one of the key physiological features includes:

[0015] Based on each key physiological feature and its corresponding classification threshold and retention duration, a corresponding physiological event is determined; the feature analysis results include at least one physiological event.

[0016] In some embodiments, the physiological events include at least one of the following: baseline events, rising events, number of rising events, falling events, short-term baseline variability events, and baseline variability events of the monitored indicators; the key physiological characteristics include characteristic values ​​of the monitored indicators.

[0017] The step of determining the corresponding physiological event based on each key physiological feature, its corresponding classification threshold, and its corresponding retention duration includes:

[0018] The baseline level is obtained by fitting the characteristic values ​​of the monitoring indicators at multiple time points. The baseline event is determined based on the baseline level, the preset baseline threshold, and the preset baseline retention time. The types of the baseline event include baseline abnormality and baseline normality.

[0019] And / or, the rise event is determined based on the monitoring indicator characteristic value, the baseline level, the rise threshold, and the preset rise duration;

[0020] And / or, the descent event is determined based on the characteristic value of the monitoring indicator, the baseline level, the descent threshold, and the preset descent duration;

[0021] And / or, count the number of rising events within a preset time period to obtain the number of rising events;

[0022] And / or, the short-term variation event is determined based on the characteristic value of the monitoring indicator, the baseline level, the preset amplitude fluctuation range, and the preset fluctuation duration;

[0023] And / or, the baseline variation event is determined based on the baseline level, the preset baseline fluctuation range, and the preset baseline retention duration.

[0024] In some embodiments, the method further includes:

[0025] Before generating the primary report, the feature analysis results are automatically scored and classified using a preset rule engine based on the risk label of the monitored object and at least one physiological event included in the feature analysis results, to obtain the target score and target level.

[0026] The process of generating a preliminary report based on the feature analysis results and the obtained comprehensive information about the monitored objects includes:

[0027] Based on the comprehensive information of the monitored object, including the target score and the target level, and at least one physiological event included in the feature analysis results, a preliminary report is generated according to a preset template; wherein, the preset template includes monitoring results and result classification.

[0028] In some embodiments, the step of automatically scoring and classifying the feature analysis results based on the risk label of the monitored object and at least one physiological event included in the feature analysis results, using a preset rule engine, to obtain a target score and target level, includes:

[0029] The feature analysis results include the types and categories of physiological events, which determine the initial score;

[0030] A safety factor is determined based on the risk label of the monitored object; the initial score is weighted using the safety factor to obtain the target score;

[0031] The target level is determined based on the target score and multiple preset level thresholds.

[0032] In some embodiments, the preset template includes monitoring results and result classification;

[0033] The step involves generating a preliminary report based on a preset template, using the comprehensive information of the monitored object, including the target score and target level, and at least one physiological event included in the feature analysis results. The report includes:

[0034] The monitoring results are generated based on at least one physiological event included in the feature analysis results; the monitoring results include reactive, non-reactive, and suspicious types;

[0035] Based on the target level, an AI model is used to generate medical order information, and the result classification is generated by combining the medical order information; the result classification includes normal, medical treatment, and retest.

[0036] In some embodiments, the method includes at least one of the following:

[0037] First item:

[0038] The construction of structured context information based on the primary report and at least one of the risk warning information includes:

[0039] Structured contextual information is constructed based on at least one physiological event, the primary report, at least one risk warning message, and basic information of the monitored object, as included in the feature analysis results.

[0040] The second item: The input of the structured context information into the retrieval-enhanced large language model to generate a monitoring report conclusion in natural language form includes:

[0041] The structured context information is identified using the retrieval-enhanced large language model, and a report content described in natural language is generated. The report content is then standardized and adjusted according to the built-in standard descriptive terminology library to generate a monitoring report conclusion in natural language form.

[0042] In some embodiments, the method further includes at least one of the following:

[0043] The first item: The report content of recognizing the structured context information and generating a natural language description using the retrieval-enhanced large language model includes:

[0044] Generate AI suggestions based on at least one of the aforementioned risk warnings;

[0045] Detailed analysis is generated based on at least one physiological event, outcome classification, and monitoring results;

[0046] The monitoring report conclusions are generated based on AI suggestions, result classification, monitoring results, and detailed analysis.

[0047] Second item:

[0048] During the process of matching each piece of information in the primary report with a preset expert experience rule base, the matching process and matching results are recorded to obtain an inference chain. The inference chain is then associated with the corresponding risk warning information to interpret at least one of the risk warning information.

[0049] Secondly, embodiments of this application provide a terminal device, the terminal device including a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement a monitoring report generation method provided in the first aspect of this application.

[0050] The embodiments of this application have the following beneficial effects:

[0051] This application inputs monitoring data sequences into a hybrid neural network model to extract local morphological features, thereby identifying key physiological characteristics and obtaining feature analysis results. Based on the feature analysis results and the comprehensive information of the monitored objects, a preliminary report is generated. Each piece of information in the preliminary report is matched with a preset expert experience rule base to generate at least one risk warning message. Structured context information is constructed based on the preliminary report and at least one risk warning message. The structured context information is input into a retrieval-enhanced large language model to generate a monitoring report conclusion in natural language form. This application considers factors such as feature analysis results, expert experience, and risk warning information, resulting in a more accurate monitoring report conclusion. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart of a monitoring report generation method according to an embodiment of this application is shown;

[0054] Figure 2 This paper illustrates an analysis flowchart of the feature analysis results in the monitoring report generation method of this application embodiment;

[0055] Figure 3 This paper illustrates a flowchart of a method for determining physiological events in the monitoring report generation method according to an embodiment of this application.

[0056] Figure 4 This paper illustrates a flowchart of a primary report generation method in the monitoring report generation method of this application.

[0057] Figure 5 This paper illustrates a flowchart of a monitoring report generation method according to an embodiment of the present application, which generates monitoring report conclusions.

[0058] Figure 6 This illustration shows a schematic diagram of a monitoring report generated using an embodiment of the present application.

[0059] Figure 7 A schematic diagram of a monitoring report generation device according to an embodiment of this application is shown.

[0060] Explanation of key component symbols:

[0061] 610 - Feature Analysis Module; 620 - Preliminary Report Generation Module; 630 - Risk Tag Generation Module; 640 - Context Construction Module; 650 - Conclusion Generation Module. Detailed Implementation

[0062] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0063] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0064] In the following text, the terms "comprising," "having," and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more combinations thereof. Furthermore, the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0065] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0066] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0067] The following examples illustrate the method for generating this monitoring report.

[0068] Figure 1 A flowchart illustrating a monitoring report generation method according to an embodiment of this application is shown. Exemplarily, the monitoring report generation method includes the following steps:

[0069] S100 inputs the monitoring data sequence into a hybrid neural network model to extract local morphological features, thereby identifying key physiological features and obtaining feature analysis results.

[0070] Monitoring data sequences are time-series data obtained by monitoring preset indicators of the monitored subject. Monitoring data sequences include, but are not limited to, fetal heart rate monitoring data sequences. For example, fetal heart rate monitoring data sequences include fetal heart rate data sequences and uterine contraction pressure data sequences. The fetal heart rate data sequence is obtained based on the fetal heart rate (FHR) curve over time, and the uterine contraction pressure data sequence is obtained based on the uterine contraction pressure curve.

[0071] If the acquired monitoring data is in the form of images, this embodiment of the application requires image processing to convert it into an initial monitoring data sequence, and also requires preprocessing the initial monitoring data sequence to obtain the monitoring data sequence. If the monitoring data sequence is a fetal heart rate monitoring data sequence, the preprocessing includes, but is not limited to, filtering and denoising, and standardizing the heart rate signal.

[0072] Local morphological features are the patterns or segments of change in monitored indicators within a short time window. Key physiological features include, for example, fetal heart rate and uterine contraction pressure. The hybrid network model includes at least two neural networks for extracting local morphological features and identifying key physiological features.

[0073] Furthermore, to improve accuracy, the monitoring data sequence includes a first type of data sequence and a second type of data sequence;

[0074] like Figure 2 As shown, the step of inputting the monitoring data sequence into a hybrid neural network model to extract local morphological features, and then identifying at least one key physiological feature, includes:

[0075] S110, the first type of data sequence and the second type of data sequence are input into the first neural network to identify local morphological features.

[0076] Exemplary, the monitoring data sequence is a fetal heart rate monitoring data sequence, the first type of data sequence is a fetal heart rate data sequence, the second type of data sequence is a uterine contraction pressure data sequence, and the first neural network includes, but is not limited to, using a convolutional neural network (CNN). Understandably, the fetal heart rate data sequence and the uterine contraction pressure data sequence are input into the convolutional neural network to identify local morphological features of the fetal heart rate curve.

[0077] S120, the local morphological features are input into the second neural network to track the changes in the local morphological features, identify at least one of the key physiological features, and generate the feature analysis results based on at least one of the key physiological features.

[0078] The second neural network includes, but is not limited to, recurrent neural networks (RNNs). RNNs are used to capture time-series patterns. CNNs can identify the shape of curves within short time windows, such as segments that rise rapidly (potential acceleration of fetal heart rate) or fall rapidly (deceleration of fetal heart rate). RNNs are responsible for tracking changes in these local morphological features, i.e., combining them with long-term context to determine whether these segments meet the duration requirements.

[0079] The second neural network is configured with multiple channels, which can output a variety of key physiological features based on different classification thresholds.

[0080] For example, a fetal heart rate data sequence of a certain length (e.g., 30 minutes) is input into a CNN. The CNN extracts the change features every minute, which are accumulated in the hidden state of the RNN. Finally, a series of key physiological features are output, such as whether there are baseline deviations, acceleration events, deceleration events, etc.

[0081] Further, in step S120, generating the feature analysis result based on at least one of the key physiological features includes:

[0082] Based on each key physiological feature and its corresponding classification threshold and retention duration, a corresponding physiological event is determined; the feature analysis results include at least one physiological event.

[0083] CNN and RNN incorporate standards from international obstetric fetal monitoring guidelines during training, making the output feature analysis results interpretable.

[0084] Exemplary, the physiological events include baseline events, rising events, number of rising events, falling events, short-term baseline variability events, and baseline variability events for the monitored indicators; the key physiological characteristics include characteristic values ​​of the monitored indicators.

[0085] like Figure 3 As shown, determining the corresponding physiological event based on each key physiological feature, its corresponding classification threshold, and its corresponding retention duration includes:

[0086] S121, a baseline level is obtained by fitting the characteristic values ​​of the monitoring indicators at multiple time points. The baseline event is then determined based on the baseline level, a preset baseline threshold, and a preset baseline maintenance duration. In other words, the baseline event is determined based on whether the duration for which the baseline level exceeds or falls below the preset baseline threshold meets the preset baseline maintenance duration. The types of baseline events include baseline anomaly and baseline normality.

[0087] The baseline level of a monitoring indicator is the stable average level of that indicator over a one-observation period (typically ≥10 minutes) in the absence of significant upward or downward events. For example, the baseline level of fetal heart rate is the stable average level of fetal heart rate over a one-observation period (typically ≥10 minutes) in the absence of significant acceleration or deceleration.

[0088] The baseline fitting of the monitoring indicators in this scheme is a relatively stable central value estimated by using a sliding window statistical or regression method after the AI ​​model (CNN-RNN hybrid neural network) extracts trends and removes outliers from the monitoring data sequence.

[0089] If the monitored indicator is fetal heart rate, the hybrid neural network model will fit a baseline fetal heart rate level (usually 110-160 bpm is considered normal). If the hybrid neural network model detects that the baseline level is consistently higher than 160 or lower than 110, the type of baseline event is determined to be a baseline abnormality.

[0090] S122, the rise event is determined based on the monitoring indicator characteristic value, the baseline level, the rise threshold, and the preset rise duration. In other words, the rise event is determined based on whether the magnitude and duration of the monitoring indicator characteristic value exceeding the baseline level correspond to the rise threshold and the preset rise duration.

[0091] If the monitored indicator is fetal heart rate, then if the hybrid neural network model detects that the fetal heart rate is higher than the baseline level than the rise threshold (e.g., 15 bpm) and lasts for at least 15 seconds (preset rise duration), then it is determined that a rise event (also known as an acceleration event) has occurred. The hybrid neural network model records the time index of each segment in the monitoring data sequence that meets the judgment condition of the rise event.

[0092] S123, the descent event is determined based on the monitoring indicator characteristic value, the baseline level, the descent threshold, and the preset descent duration. In other words, the descent event is determined based on whether the magnitude and duration of the monitoring indicator characteristic value falling below the baseline level correspond to the descent threshold and the preset descent duration.

[0093] If the monitored indicator is fetal heart rate, then if the fetal heart rate characteristic value is greater than or equal to the descent threshold (15 bpm) below the baseline level and continues to decrease for a preset duration (e.g., 15 seconds), the hybrid neural network model will identify a descent event (also known as a deceleration event). Types of deceleration events include early-onset, late-onset, and variable decelerations, which are further classified based on the morphology of the fetal heart rate curve, its timing, and its relationship to uterine contractions.

[0094] S124, count the number of rising events within a preset time period to obtain the number of rising events.

[0095] If the monitored indicator is fetal heart rate, for example, counting the number of acceleration events marked by a hybrid neural network model within 30 minutes, the output would be "Number of accelerations = n times / 30 minutes". The number of rising events is also called the number of acceleration events, and the result type for the number of acceleration events includes normal and abnormal. For example, a number of acceleration events greater than or equal to 2 times / 20 minutes is considered normal.

[0096] S125, the short-term variation event is determined based on the monitoring indicator characteristic value, the baseline level, the preset amplitude fluctuation range, and the preset fluctuation duration. In other words, the short-term variation event is determined based on whether the magnitude and duration of the monitoring indicator characteristic value exceeding the baseline level correspond to the preset amplitude fluctuation range and the preset fluctuation duration. If the magnitude of the monitoring indicator characteristic value exceeding the baseline level meets the preset amplitude fluctuation range, and the duration of the monitoring indicator characteristic value exceeding the baseline level meets the preset fluctuation duration, then the short-term variation event is determined.

[0097] If the monitored indicator is fetal heart rate, then the instantaneous amplitude fluctuation range of the fetal heart rate is calculated. A fluctuation range of 5–25 bpm is considered normal. For example, regression prediction or classification is performed on the variability per minute. Types of short-term variability events include normal, too low, and too high.

[0098] S126, the baseline variation event is determined based on the baseline level, the preset baseline fluctuation range, and the preset baseline maintenance duration. In other words, the baseline variation event is determined based on whether the duration for which the baseline level exceeds the preset baseline fluctuation range meets the preset baseline maintenance duration. If the duration for which the baseline level exceeds the preset baseline fluctuation range meets the preset baseline maintenance duration, then the baseline variation event is determined.

[0099] If the monitored indicator is fetal heart rate, then the fluctuation range of the fetal heart rate baseline level, i.e., the range of short-term heart rate fluctuations above and below the baseline level (preset baseline fluctuation range), is considered. Under normal circumstances, the short-term fluctuation range of fetal heart rate is between 5 and 25 bpm. A hybrid neural network model extracts curve details through a CNN and continuously tracks changes through an RNN to calculate the short-term fluctuation range and determine whether it is within the normal range (preset baseline fluctuation range). Here, the short-term fluctuation range is the change in baseline level.

[0100] S200: Based on the feature analysis results and the comprehensive information of the monitored objects obtained, a preliminary report is generated.

[0101] Feature analysis results are clinically significant physiological events that are automatically extracted and identified from the original monitoring data sequences using hybrid neural network models (such as CNN+RNN).

[0102] Comprehensive information on monitored subjects refers to background information and risk assessment parameters related to the individual being monitored, in addition to physiological signals, which are used to assist in report generation and risk assessment.

[0103] The preliminary report is a structured intermediate report generated according to a preset template based on the feature analysis results and comprehensive information of the monitored objects. It serves as the basis for subsequent risk warning matching and the generation of final conclusions.

[0104] S300, each piece of information in the primary report is matched with a preset expert experience rule base to generate at least one risk warning message to achieve knowledge enhancement.

[0105] The risk warning information includes causal factors. The preset expert experience rule base is constructed based on industry standards for monitoring indicators, expert experience, and AI large-scale models. If the monitoring data sequence is a fetal heart rate monitoring data sequence, the preset expert experience rule base is constructed based on fetal heart rate monitoring industry standards, expert experience, and AI large-scale models.

[0106] Based on obstetric fetal monitoring guidelines and the experience of senior obstetricians, and combined with general knowledge retrieved by AI large-scale models (such as deepseekR1), a preset expert experience rule base is generated. This preset expert experience rule base includes thousands of risk rules (or experience rules), and each piece of information in the primary report is matched with each risk rule in the preset expert experience rule base. For example, "repeated late decelerations and gestational age greater than 41 weeks" corresponds to "hypoxia risk," "baseline level >170 with reduced variability" corresponds to "possible fetal distress," and "no accelerations for 40 minutes" corresponds to "unresponsive, requiring retesting," and so on. It is evident that each rule in the preset expert experience rule base includes triggering conditions and corresponding risk labels. Triggering conditions include information from the primary report and gestational age. Risk labels refer to a type of structured metadata labeled for individual monitoring subjects, used to characterize potential medical risk factors during their pregnancy, including but not limited to: abnormal gestational age (such as post-term pregnancy), pregnancy complications (such as gestational diabetes mellitus, preeclampsia), and a history of adverse pregnancy outcomes. The information in the primary report (including baseline events, acceleration events, deceleration events, etc.) is matched against trigger conditions in a pre-defined expert experience rule base. If the information in the primary report meets the trigger conditions of a certain rule, a corresponding risk warning label (risk warning information) is added to the result. For example, if a "continuous variable deceleration" event is detected and the pregnant woman is in a high-risk gestational age, the "possible fetal hypoxia" label is automatically added.

[0107] In one implementation, during the matching process of each piece of information in the primary report with a preset expert experience rule base, the matching process and results are recorded to obtain an inference chain. This inference chain is then associated with the corresponding risk warning information to explain at least one risk warning message. This explains why the AI ​​model arrives at its conclusion, ensuring the transparency and credibility of the method in this application's embodiments, and also facilitating later review or doctor's viewing and optimization of the auxiliary pathway.

[0108] With knowledge enhancement, AI no longer just provides surface-level indicators but can also offer interpretations with certain clinical significance, laying the foundation for subsequent report generation. Compared to the black-box output of pure machine learning, this reasoning that integrates expert experience and knowledge improves the interpretability of the results and the ability to make independent judgments that conform to medical logic.

[0109] This application's embodiments can complete the preliminary interpretation and classification of monitoring reports without human intervention, saving manpower for subsequent processing. Several months of testing showed that for clearly normal reports, the AI ​​interpretation matched human conclusions with a 98% accuracy rate, and it can handle approximately 40% of low-risk reports, significantly improving work efficiency.

[0110] S400, construct structured context information based on the primary report and at least one of the risk warning information.

[0111] If the monitored indicator is fetal heart rate, then the risk warning information includes the risk of hypoxia, and the need for retesting if there is no response.

[0112] S500, the structured context information is input into the retrieval-enhanced large language model to generate a monitoring report conclusion in natural language form. In this embodiment, the monitoring indicator is fetal heart rate, and the generated monitoring report conclusion is the content of the fetal heart rate monitoring report.

[0113] Structured contextual information is converted into prompt words, which are then input into the retrieval-enhanced large language model.

[0114] Retrieval-Augmented Generation (RAG) and Large Language Model (LLM) are among the technologies employed in large language models, including but not limited to RAG+LLM. RAG+LLM is a technical architecture that combines Retrieval-Augmented Generation (RAG) with Large Language Model (LLM), and is currently an advanced paradigm applied in artificial intelligence, natural language processing, and knowledge-intensive tasks.

[0115] In one embodiment, the method further includes:

[0116] S130, before generating the primary report, based on the risk label of the monitored object and at least one physiological event included in the feature analysis results, the feature analysis results are automatically scored and classified using a preset rule engine to obtain the target score and target level.

[0117] The preset rule engine includes, but is not limited to, the Fischer score, which is commonly used in clinical practice.

[0118] Exemplary, in step S130, the step of automatically scoring and classifying the feature analysis results based on the risk label of the monitored object and at least one physiological event included in the feature analysis results, using a preset rule engine, to obtain a target score and target level, includes:

[0119] S131, the feature analysis results include the types and categories of physiological events, and the initial score is determined.

[0120] The physiological events include baseline events, rising events (acceleration events), number of rising events, falling events (deceleration events), short-term baseline variation events, and baseline variation events for the monitored indicators; the key physiological characteristics include the characteristic values ​​of the monitored indicators.

[0121] For example, scoring based on a custom 9-point scale:

[0122] If the baseline event type is baseline normal (baseline event is not identified as abnormal), the baseline short-term variation event type is variation normal, the acceleration event type is normal, and there is no deceleration event, then the target score is a maximum of 9 points.

[0123] If there are suspicious indicators (such as the type of baseline short-term variation event being too small or the type of deceleration event being occasional deceleration), the target score is 7-8 points.

[0124] Occasional decelerations refer to non-repetitive, isolated, or sporadic decreases in fetal heart rate during fetal heart rate monitoring. They are usually associated with uterine contractions but do not exhibit a regular pattern. Types of deceleration events include early decelerations, variable decelerations, and late decelerations. If any type of deceleration event occurs only 1–2 times and does not meet the "repetitive" criterion, it is marked as an "occasional deceleration."

[0125] If a significant anomaly occurs (the type of deceleration event is continuous deceleration or the type of baseline event is baseline anomaly), the target score is less than or equal to 6 points, and it is judged as high risk.

[0126] Sustained deceleration is defined as a decrease in fetal heart rate of ≥15 bpm from baseline, lasting ≥2 minutes but <10 minutes. If it exceeds 10 minutes, it is considered a new baseline change. This is an event-related abnormal waveform, suggesting potential risks such as hypoxia or umbilical cord compression.

[0127] S132, determine the corresponding safety factor based on the risk label of the monitored object; use the safety factor to weight the initial score to obtain the target score.

[0128] If the monitored indicator is fetal heart rate, then in this embodiment, the score is weighted in conjunction with the pregnant woman's basic information. For example, if the risk label is high-risk gestational age (post-term pregnancy, etc.), patient with complications, or user comments (abnormal fetal movement count), a slightly lower safety factor is given even if the fetal monitoring indicators are normal, prompting further attention. However, the number of risk labels does not directly add up the score; instead, the safety factor is determined by the preset high-risk type labels.

[0129] S133, determine the target level based on the target score and multiple preset level threshold values.

[0130] For example, based on the target score, the initial report is automatically divided into three categories: normal, suspicious, and high-risk. A score of 9 is defined as normal; a score of 7-8 is defined as suspicious and requires retesting; a score of 6 or below is defined as high-risk, and it is recommended that the monitored individual seek medical attention.

[0131] Further, in step S200, generating a preliminary report based on the feature analysis results and the obtained comprehensive information of the monitored objects includes:

[0132] S210, based on the comprehensive information including basic information, the target score and the target level, and at least one physiological event included in the feature analysis results, a preliminary report is generated according to a preset template; wherein, the preset template includes monitoring results and result classification; the preliminary report includes: monitoring results and result classification. Monitoring results include reactive / non-reactive types, and result classification includes normal, medical attention required, and retest.

[0133] Furthermore, such as Figure 4 As shown, the step of generating a preliminary report according to a preset template, based on the comprehensive information of the monitored object including the target score and the target level, and at least one physiological event included in the feature analysis results, includes:

[0134] S211, the monitoring results are generated based on at least one physiological event included in the feature analysis results; the monitoring results include reactive, non-reactive, and questionable types. For example, if there is an acceleration event but no deceleration event, it is determined to be a "reactive" NST. NST (Non-Stress Test) is a clinically common type of fetal heart rate monitoring used to determine whether the fetus is reactive.

[0135] S212, Based on the target level, an AI model is used to generate medical order information, and the result classification is generated by combining the medical order information; the result classification includes normal, medical treatment, and retest. For example, if the corresponding score is normal, it is marked as "normal" and the generated medical order information is "routine follow-up is required"; if the score is low, the generated medical order information is "medical treatment is required".

[0136] The primary report includes: physiological events, outcome classification, monitoring results, and at least one risk warning message.

[0137] In one implementation, step S400, constructing structured context information based on the primary report and at least one of the risk warning information, includes:

[0138] S410, construct structured context information based on at least one physiological event (fetal monitoring analysis result) included in the feature analysis results, the primary report, at least one risk warning message, and basic information of the monitored object.

[0139] Key input elements are aggregated and fed into the retrieval-enhanced large language model. Input data includes: basic information of the monitored objects, feature analysis results, preliminary reports, and matched risk warning information (such as hypoxia risk, need for re-examination, etc.). This data forms a structured contextual information, which is then fed into the LLM through prompt words.

[0140] Basic information about the pregnant woman: age, gestational week, risk factors, remarks, etc.

[0141] Feature analysis results: baseline level, baseline variation events, number of acceleration events, deceleration event details, etc.

[0142] Primary report: such as monitoring results (responsive / non-responsive) and result classification (normal / retest / medical visit).

[0143] In one implementation, such as Figure 5 As shown, in step S500, the step of inputting the structured context information into the retrieval-enhanced large language model to generate a monitoring report conclusion in natural language form includes:

[0144] S510, the structured context information is identified using the retrieval-enhanced large language model and a report content described in natural language is generated. The report content is then standardized and adjusted according to the built-in standard descriptive terminology library to generate a monitoring report conclusion in natural language form.

[0145] As an example, the retrieval-enhanced large language model is constructed using the RAG+LLM technical framework. LLM (Large Language Model) refers to a deep learning model trained on massive amounts of text data (such as GPT, BERT, ChatGLM, and Tongyi Qianwen), possessing powerful language understanding and generation capabilities. It can answer questions, write, and reason in a "human-like" manner, but its knowledge comes from the training data, leading to problems such as knowledge rigidity, difficulty in updating, and susceptibility to illusions. RAG (Retrieval-Enhanced Generation) is a method to overcome the limitations of LLM. It introduces an external knowledge base to "search for relevant information" before generating an answer, allowing the LLM to respond based on the found content.

[0146] Understandably,

[0147] The LLM is used to identify the structured context information and generate a report content described in natural language. The report content is then retrieved through RAG and standardized according to the built-in standard descriptive terminology library to generate a monitoring report conclusion in natural language form.

[0148] The standard description terminology library includes relevant medical knowledge snippets, standard descriptions from fetal heart rate monitoring guidelines, and built-in explanation templates. For example, to ensure the accuracy of the LLM-generated content, the system uses the RAG to retrieve relevant medical knowledge snippets for reference, such as extracting standard descriptions of normal and abnormal conditions from fetal heart rate monitoring guidelines, or referencing standard descriptions from the built-in explanation template library. This allows the system to refer to authoritative statements when generating monitoring report conclusions, reducing inaccuracies caused by arbitrary interpretation.

[0149] Furthermore, such as Figure 5 As shown, in step S510, the step of using the retrieval-enhanced large language model to identify the structured context information and generate a report content with a natural language description includes:

[0150] S511, Generate AI suggestions based on the at least one risk warning information.

[0151] AI suggestions are used to provide advice to the monitored subjects, such as "It is recommended that pregnant women count fetal movements daily, and each monitoring session should last no less than 40 minutes in the later stages." AI suggestions are derived from pre-set rules (such as general recommendations for normal reporting) and LLM's consideration of risk warning information. If there are high-risk factors, specific precautions may be added.

[0152] S512 generates a detailed analysis based on at least one physiological event, outcome classification, and monitoring results.

[0153] Result classification is used to give a simple and clear conclusion, such as "Result classification: Normal". Monitoring results are used to correspond to the NST conclusion, such as "Monitoring results: Reactive".

[0154] Detailed analysis (also known as AI analysis) includes: explaining various physiological events in natural language, such as "the fetal heart rate baseline is within the normal range (150 bpm), the variability is normal (10 bpm), there are 6 acceleration events and no deceleration events within 30 minutes, all core indicators meet the standards, and the overall assessment is normal." The content retrieval enhancement big language model of detailed analysis will automatically fill in the explanation based on the input indicators.

[0155] S513 generates a monitoring report conclusion based on AI suggestions, result classification, monitoring results (also known as AI monitoring results), and detailed analysis.

[0156] For example, the monitoring report conclusion includes: "Monitoring results: reactive, classification: normal. Detailed analysis: The baseline fetal heart rate monitoring was approximately 150 bpm, within the normal range (110~160 bpm); the fetal heart rate amplitude variation was approximately 10 bpm, which is within the normal range; there were 6 heart rate acceleration events within 30 minutes (meeting the standard of ≥2 times / 20 minutes), and no abnormal decelerations were observed. All indicators show that the fetal heart rate monitoring results are normal. It is recommended that the pregnant woman continue to monitor fetal movement daily and have regular follow-up fetal monitoring, with each monitoring session lasting no less than 40 minutes."

[0157] This application embodiment also includes: verifying the monitoring report conclusions against key values ​​within the structured context information, and after confirming that there are no errors, generating the final monitoring report text and submitting it to the doctor's workbench for review. Figure 6 As shown, in a human-machine collaboration mode, doctors can quickly confirm or modify minor text based on the final monitoring report text before issuing it. In any case, the AI-generated final monitoring report text significantly improves report interpretation efficiency and, under most normal circumstances, can be directly used for feedback to users.

[0158] The application of RAG+LLM enables this application to automatically generate professional reports, integrating complex signal processing and rule-based reasoning results into natural language. This not only reflects the deep involvement of AI but also ensures the readability and accuracy of the output.

[0159] Figure 7 A schematic diagram of a monitoring report generation device according to an embodiment of this application is shown. Exemplarily, the monitoring report generation device includes: a feature analysis module 610, a primary report generation module 620, a risk label generation module 630, a context construction module 640, and a conclusion generation module 650.

[0160] The feature analysis module 610 is used to input the monitoring data sequence into the hybrid neural network model to extract local morphological features, thereby identifying key physiological features and obtaining feature analysis results.

[0161] The primary report generation module 620 generates a primary report based on the feature analysis results and the comprehensive information of the monitored objects obtained.

[0162] The risk label generation module 630 is used to match each piece of information in the primary report with a preset expert experience rule base to generate at least one risk warning message;

[0163] Context building module 640 is used to build structured context information based on the primary report and at least one of the risk warning information;

[0164] The conclusion generation module 650 is used to input the structured context information into the retrieval-enhanced large language model to generate a monitoring report conclusion in natural language form. It is understood that the apparatus in this embodiment corresponds to the monitoring report generation method of the above embodiments, and the options in the above embodiments are also applicable to this embodiment, therefore, they will not be described again here.

[0165] This application also provides a terminal device, exemplary of which includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to enable the terminal device to perform the functions of the various modules in the above-described monitoring report generation method or the above-described monitoring report generation apparatus.

[0166] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0167] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving execution instructions.

[0168] This application also provides a computer-readable storage medium for storing the computer program used in the aforementioned terminal device. For example, the computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0169] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0170] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0171] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0172] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for generating a monitoring report, characterized in that, include: The monitoring data sequence is input into a hybrid neural network model to extract local morphological features, thereby identifying key physiological characteristics. The feature analysis results were obtained; Based on the feature analysis results and the comprehensive information of the monitored objects obtained, a preliminary report is generated; Each piece of information in the primary report is matched with a preset expert experience rule base to generate at least one risk warning message; Construct structured contextual information based on the primary report and at least one of the risk warning messages; The structured context information is input into the retrieval-enhanced large language model to generate a monitoring report conclusion in natural language form; Before generating the primary report, the feature analysis results are automatically scored and classified using a preset rule engine based on the risk label of the monitored object and at least one physiological event included in the feature analysis results, to obtain the target score and target level. The step of generating a preliminary report based on the feature analysis results and the acquired comprehensive information of the monitored objects includes: generating the preliminary report according to a preset template based on the target score and target level included in the comprehensive information of the monitored objects and at least one physiological event included in the feature analysis results; wherein the preset template includes monitoring results and result classification.

2. The monitoring report generation method according to claim 1, characterized in that, The monitoring data sequence includes a first type of data sequence and a second type of data sequence; The monitoring data sequence is input into a hybrid neural network model to extract local morphological features, thereby identifying key physiological characteristics. The feature analysis results obtained include: The first type of data sequence and the second type of data sequence are input into the first neural network to identify the local morphological features; The local morphological features are input into a second neural network to track changes in the local morphological features, identify at least one of the key physiological features, and generate the feature analysis results based on at least one of the key physiological features.

3. The monitoring report generation method according to claim 2, characterized in that, The generation of the feature analysis result based on at least one of the key physiological features includes: Based on each key physiological feature and its corresponding classification threshold and retention duration, a corresponding physiological event is determined; the feature analysis results include at least one physiological event.

4. The monitoring report generation method according to claim 3, characterized in that, The physiological events include at least one of the following: baseline events, rising events, number of rising events, falling events, short-term baseline variability events, and baseline variability events of the monitored indicators; the key physiological characteristics include the characteristic values ​​of the monitored indicators. The step of determining the corresponding physiological event based on each key physiological feature, its corresponding classification threshold, and its corresponding retention duration includes: The baseline level is obtained by fitting the characteristic values ​​of the monitoring indicators at multiple time points. The baseline event is determined based on the baseline level, the preset baseline threshold, and the preset baseline retention time. The types of the baseline event include baseline abnormality and baseline normality. And / or, the rise event is determined based on the monitoring indicator characteristic value, the baseline level, the rise threshold, and the preset rise duration; And / or, the descent event is determined based on the characteristic value of the monitoring indicator, the baseline level, the descent threshold, and the preset descent duration; And / or, count the number of rising events within a preset time period to obtain the number of rising events; And / or, the short-term variation event is determined based on the characteristic value of the monitoring indicator, the baseline level, the preset amplitude fluctuation range, and the preset fluctuation duration; And / or, the baseline variation event is determined based on the baseline level, the preset baseline fluctuation range, and the preset baseline retention duration.

5. The monitoring report generation method according to claim 1, characterized in that, The step of automatically scoring and classifying the feature analysis results based on the risk label of the monitored object and at least one physiological event included in the feature analysis results, using a preset rule engine, to obtain a target score and target level, includes: The feature analysis results include the types and categories of physiological events, which determine the initial score; A safety factor is determined based on the risk label of the monitored object; the initial score is weighted using the safety factor to obtain the target score; The target level is determined based on the target score and multiple preset level thresholds.

6. The monitoring report generation method according to claim 1, characterized in that, The preset template includes monitoring results and result classification; The step involves generating a preliminary report based on a preset template, using the comprehensive information of the monitored object, including the target score and target level, and at least one physiological event included in the feature analysis results. The report includes: The monitoring results are generated based on at least one physiological event included in the feature analysis results; the monitoring results include reactive, non-reactive, and suspicious types; Based on the target level, an AI model is used to generate medical order information, and the result classification is generated by combining the medical order information; the result classification includes normal, medical treatment, and retest.

7. The monitoring report generation method according to claim 1, characterized in that, The method includes at least one of the following: First item: The construction of structured context information based on the primary report and at least one of the risk warning information includes: Structured contextual information is constructed based on at least one physiological event, the primary report, at least one risk warning message, and basic information of the monitored object, as included in the feature analysis results. The second item: The input of the structured context information into the retrieval-enhanced large language model to generate a monitoring report conclusion in natural language form includes: The structured context information is identified using the retrieval-enhanced large language model, and a report content described in natural language is generated. The report content is then standardized and adjusted according to the built-in standard descriptive terminology library to generate a monitoring report conclusion in natural language form.

8. The monitoring report generation method according to claim 7, characterized in that, The method further includes at least one of the following: The first item: The report content of recognizing the structured context information and generating a natural language description using the retrieval-enhanced large language model includes: Generate AI suggestions based on at least one of the aforementioned risk warnings; Detailed analysis is generated based on at least one physiological event, outcome classification, and monitoring results; The monitoring report conclusions are generated based on AI suggestions, result classification, monitoring results, and detailed analysis. Second item: During the process of matching each piece of information in the primary report with a preset expert experience rule base, the matching process and matching results are recorded to obtain an inference chain. The inference chain is then associated with the corresponding risk warning information to interpret at least one of the risk warning information.

9. A terminal device, characterized in that, The terminal device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the monitoring report generation method according to any one of claims 1-8.

Images