Perinatal period grading management system based on AI risk assessment

Abstract

The invention relates to the technical field of medical health information, in particular to a perinatal period grading management system based on AI risk assessment, which comprises a data acquisition layer, a feature engineering layer, an AI risk assessment module, a five-color dynamic grading module and an intervention push management module. According to the scheme, a multi-scale decision table is constructed by adopting a multi-scale data fusion method, modeling is carried out on data of different sources and different scales, information gain is introduced as a scale selection standard, a high-fidelity fusion feature vector is formed, discriminative information of each data source under the optimal scale is reserved, and a high-fidelity fusion feature vector is formed. High-quality input with a clear structure and a uniform scale is provided for a subsequent AI risk assessment module; an improved Transform model is designed, a time dependence matrix and a dynamic weight adjustment mechanism are embedded, deep modeling of time series data and generation of interpretable dynamic rules are achieved, online self-updating of the rules is supported, and the accuracy and real-time performance of risk assessment are improved.

Description

Technical Field

[0001] This invention relates to the field of medical and health information technology, and in particular to a perinatal graded management system based on AI risk assessment, which is applicable to high-risk screening, dynamic early warning and intervention support for pregnant women from preconception to 42 days postpartum. Background Technology

[0002] Perinatal management refers to the systematic health monitoring and intervention of pregnant women and newborns during the critical period from 28 weeks of gestation to 1 week postpartum. Its core objectives are to identify high-risk factors early, prevent and reduce maternal and neonatal complications and mortality, ensure maternal and infant safety, and improve the quality of the newborn population. It is an important component of the maternal and child health care system. Currently, perinatal management in my country still faces multiple challenges: the identification of high-risk pregnant women relies heavily on the subjective experience of clinicians, and different physicians interpret the same indicators (such as blood pressure, fetal heart rate variability, and ultrasound parameters) significantly differently. At the same time, there is a serious disconnect between in-hospital and out-of-hospital data. Dynamic health data collected at home (such as remote fetal heart rate monitoring, blood pressure, blood glucose, and fetal movement records) is difficult to exchange and integrate with hospital information systems such as HIS, LIS, and PACS in real time. More importantly, various types of data are highly heterogeneous in terms of time granularity, sampling frequency, and semantic structure, making it difficult for traditional systems to effectively integrate them. In addition, existing risk assessments are mostly based on static thresholds or fixed rules (such as "blood pressure ≥140 / 90 is high-risk"), which cannot capture the dynamic evolution of physiological parameters during pregnancy, lack individualized and adaptive modeling capabilities, and make it difficult to achieve accurate and real-time risk warnings and graded interventions. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a perinatal triage management system based on AI risk assessment. Addressing the problems of fragmented data between hospitals and other institutions, difficulties in interoperability between home-based dynamic monitoring and hospital systems, and the high heterogeneity of multi-source data in terms of time granularity, sampling frequency, and semantic structure, traditional systems lack effective fusion and dynamic modeling capabilities, this solution employs a multi-scale data fusion method to construct a multi-scale decision table. It models data from different sources at different scales, introduces information gain as a scale selection criterion, and automatically identifies the scale at which each feature has the strongest discriminative power for risk prediction, forming a high-fidelity fused feature vector. This retains the discriminative information of each data source at its optimal scale, providing a high-quality input with a clear structure and unified scale for the subsequent AI risk assessment module. Furthermore, addressing the problem that existing risk assessments are mostly based on static thresholds or fixed rules, failing to capture the dynamic evolution of physiological parameters during pregnancy and lacking individualized and adaptive modeling capabilities, this invention designs an improved Transformer model, embedding a time-dependent matrix and a dynamic weight adjustment mechanism. This enables deep modeling of time-series data and interpretable dynamic rule generation, and supports online self-updating of rules, improving the accuracy and real-time performance of risk assessment.

[0004] This invention provides a perinatal tiered management system based on AI risk assessment. The system includes a data acquisition layer, a feature engineering layer, an AI risk assessment module, a five-color dynamic tiering module, and an intervention push management module.

[0005] The data acquisition layer connects to the hospital information system, smart wearable devices, home medical devices, mobile apps and mini-programs to collect perinatal data of pregnant women from multiple sources. The perinatal data is then layered according to different time scales to obtain multi-scale perinatal data.

[0006] The feature engineering layer cleans, formats, and standardizes the collected multi-scale perinatal data, and uses a multi-scale data fusion method to fuse the multi-scale perinatal data to obtain a multi-scale feature set.

[0007] The AI ​​risk assessment module uses an improved Transformer model to capture time-series features in a multi-scale feature set, generate dynamic rules, and output risk probabilities and rule interpretations.

[0008] The five-color dynamic grading module automatically maps risk probabilities to five levels of risk labels: green, yellow, orange, red, and purple, and dynamically updates the risk probabilities output by the AI ​​risk assessment module.

[0009] The intervention push management module establishes electronic cases for high-risk pregnant women with orange and red risk labels, automatically triggering SMS reminders, medical assistant group notifications, and referral suggestions.

[0010] Furthermore, the feature engineering layer uses a multi-scale data fusion method to fuse multi-scale perinatal data. The multi-scale data fusion method specifically includes the following steps:

[0011] Step S1: Decision table construction. Construct a multi-scale decision table for each type of multi-scale perinatal data. The multi-scale decision table includes an object set, a condition attribute set, and a decision attribute. All multi-scale decision tables share the same set of objects and assign a uniform decision attribute to each object.

[0012] Step S2: Scale partitioning, divide each conditional attribute in the conditional attribute set into N scales according to time from fine to coarse, and construct a scale tree;

[0013] Step S3: Information gain calculation. Calculate the information gain of each conditional attribute in the conditional attribute set at different scales. The formula used is as follows: ;

[0014] In the formula, Indicates the first The conditional attribute in the first Attribute representation at each scale The overall entropy measures the decision-making attribute. Uncertainty Conditional entropy measures the conditional entropy of known properties. under conditions Uncertainty Indicates information gain;

[0015] Step S4: Scale selection. For each conditional attribute, compare the information gain at different scales. If the information gain of the current conditional attribute is similar at N scales, the coarsest scale is selected first to represent the current conditional attribute. If the information gain values ​​differ greatly, the scale with the largest information gain value is selected as the representative scale of the current conditional attribute.

[0016] Step S5: Feature fusion. Select an optimal scale as the representative scale for each conditional attribute to form a feature set. Extract the value of each conditional attribute at its optimal scale to form a fusion decision table with consistent scale, which serves as a multi-scale fusion feature set.

[0017] Step S6: Establish scale dependencies. For attributes with hierarchical dependencies, model the inclusion relationship through the scale tree of the multi-scale decision table. Use depth-first traversal of the scale tree to compress redundant information and retain the most discriminative scale while ensuring decision consistency.

[0018] Furthermore, the AI ​​risk assessment module uses an improved Transformer model to capture time-series features in a multi-scale feature set, specifically including the following steps:

[0019] Step M1: Construct a temporal input sequence by organizing the multi-scale fusion features in the multi-scale fusion feature set of each pregnant woman into a temporal feature sequence in chronological order;

[0020] Step M2: Position encoding, using a cosine function to add position encoding to the temporal features in the temporal feature sequence, to obtain the position-embedded temporal feature sequence;

[0021] Step M3: Transformer encoder construction, using an L-layer Transformer encoder to process the temporal feature sequence of location embeddings. Each layer of the Transformer encoder contains a multi-head self-attention and feedforward network.

[0022] Step M4: Construct a time-aware dynamic self-attention mechanism, build a time dependency matrix and add it to the self-attention mechanism of the Transformer encoder, and calculate the time correlation using the following formula: ; ;

[0023] In the formula, and For indexing, Represents the first element in the time dependency matrix. line, number Column elements, and They represent the first The time point and the The actual timestamp of each point in time. The time decay coefficient, It is an exponential function. This is the attention weight matrix. For querying the matrix, The key matrix, The time dependency matrix, Let be the dimension of the key matrix. For normalization function, This is a transpose operation;

[0024] Step M5: Dynamic weight adjustment mechanism. After each Transformer layer, a gated dynamic weight module is introduced to adjust the weights of the Transformer encoder, resulting in an improved Transformer model. This includes the following steps:

[0025] Step M51: Global state representation construction. Calculate the global context summary for the hidden sequence output by the current layer Transformer encoder to obtain the global vector. The hidden sequence is the output of the Transformer encoder except for the last layer Transformer encoder.

[0026] Step M52: Dynamic weight vector generation, using a lightweight gated network to calculate the weight vector obtained by concatenating the hidden state vector and the global vector at the current time point;

[0027] Step M53: Hidden state update. Use the weight vector to perform a weighted fusion of the current hidden state vector and the hidden state vector at the previous time point to obtain a new hidden state vector.

[0028] Step M6: Rule extraction. The attention weights of the last Transformer encoder layer are used to identify the key risk time windows of the embedded temporal feature sequence, and dynamic rules are extracted by combining feature importance. The risk probability and rule interpretation are then output.

[0029] Step M7: Rule confidence and coverage evaluation. For each mined rule, calculate two metrics: rule mining accuracy and rule coverage, and optimize the improved Transformer model.

[0030] The beneficial effects achieved by the present invention using the above solution are as follows:

[0031] (1) In response to the problem that data from inside and outside the hospital are fragmented, home dynamic monitoring is difficult to communicate with the hospital system, and multi-source data are highly heterogeneous in terms of time granularity, sampling frequency and semantic structure, and traditional systems lack effective fusion and dynamic modeling capabilities, this solution adopts a multi-scale data fusion method to construct a multi-scale decision table, models data from different sources and at different scales, introduces information gain as a scale selection criterion, automatically identifies at which scale each feature has the most discriminative power for risk prediction, forms a high-fidelity fusion feature vector, retains the discriminative information of each data source at its optimal scale, and provides high-quality input with clear structure and unified scale for the subsequent AI risk assessment module;

[0032] (2) In view of the problem that existing risk assessments are mostly based on static thresholds or fixed rules, which cannot capture the dynamic evolution of physiological parameters during pregnancy and lack individualized and adaptive modeling capabilities, an improved Transformer model is designed, which embeds a time-dependent matrix and a dynamic weight adjustment mechanism to achieve deep modeling of time-series data and the generation of interpretable dynamic rules, and supports online self-updating of rules, thereby improving the accuracy and real-time performance of risk assessment, while enhancing clinical interpretability and system adaptability. The rule update cycle is shortened from several months to hours, effectively supporting individualized, dynamic and intelligent five-color grading management of the perinatal period, and providing more accurate and timely early warning and intervention for high-risk pregnant women. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of a perinatal triage management system based on AI risk assessment proposed in this invention;

[0034] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0036] Example 1, see Figure 1 The present invention provides a perinatal grading management system based on AI risk assessment. The system includes a data acquisition layer, a feature engineering layer, an AI risk assessment module, a five-color dynamic grading module, and an intervention push management module.

[0037] The data acquisition layer connects to the hospital information system, smart wearable devices, home medical devices, mobile apps and mini-programs to collect perinatal data of pregnant women from multiple sources. The perinatal data is then layered according to different time scales to obtain multi-scale perinatal data.

[0038] The feature engineering layer cleans, formats, and standardizes the collected multi-scale perinatal data, and uses a multi-scale data fusion method to fuse the multi-scale perinatal data to obtain a multi-scale feature set.

[0039] The AI ​​risk assessment module uses an improved Transformer model to capture time-series features in a multi-scale feature set, generate dynamic rules, and output risk probabilities and rule interpretations.

[0040] The five-color dynamic grading module automatically maps risk probabilities to five levels of risk labels: green, yellow, orange, red, and purple, and dynamically updates the risk probabilities output by the AI ​​risk assessment module.

[0041] The intervention push management module establishes electronic cases for high-risk pregnant women with orange and red risk labels, automatically triggering SMS reminders, medical assistant group notifications, and referral suggestions.

[0042] Example 2, based on the above examples, involves the hospital information system at the data acquisition layer interfacing with the hospital information system and electronic medical record system via the FHIR (Fast Healthcare Interoperability Resources) protocol. Images from the image archiving and communication system are used to obtain structured reports via the DICOM protocol. The information collected by the hospital information system includes the pregnant woman's basic information: age, obstetric history, past medical history, BMI, and blood type; laboratory data: complete blood count, urine protein, liver and kidney function, glucose tolerance, and thyroid function; imaging data: ultrasound report (fetal biparietal diameter, amniotic fluid index, umbilical artery S / D value), electrocardiogram, and fetal heart rate monitoring; and clinical records: chief complaint, physical signs, diagnostic conclusions, medication records, and delivery method.

[0043] The home medical devices include a home fetal heart monitor, a smart blood pressure monitor, a blood glucose meter, a weight scale, and a wearable ECG patch. They can be viewed through a mobile app and a mini-program. The mobile app and mini-program simultaneously record the pregnant woman's pregnancy time and status, and automatically synchronize prenatal check-up plans and follow-up appointment times.

[0044] Example 3, based on the above examples, describes a feature engineering layer that uses a multi-scale data fusion method to fuse multi-scale perinatal data. The multi-scale data fusion method specifically includes the following steps:

[0045] Step S1: Decision table construction. Construct a multi-scale decision table for each type of multi-scale perinatal data. The multi-scale decision table includes an object set, a condition attribute set, and a decision attribute. All multi-scale decision tables share the same set of objects and assign a uniform decision attribute to each object.

[0046] Step S2: Scale partitioning. For each conditional attribute in the set of conditional attributes, divide it into N scales according to time from fine to coarse. Construct a scale tree. The scale tree contains a root node and M leaf nodes. The root node is the attribute name, and the leaf nodes are tuples representing different scales. The scale tree structure is from top to bottom, and the scales become more refined as you go down.

[0047] Step S3: Information gain calculation. Calculate the information gain of each conditional attribute in the conditional attribute set at different scales. The formula used is as follows: ;

[0048] In the formula, Indicates the first The conditional attribute in the first Attribute representation at each scale The overall entropy measures the decision-making attribute. Uncertainty Conditional entropy measures the conditional entropy of known properties. under conditions Uncertainty Indicates information gain;

[0049] Step S4: Scale selection. For each conditional attribute, compare the information gain at different scales. If the information gain of the current conditional attribute is similar at N scales, the coarsest scale is selected first to represent the current conditional attribute. If the information gain values ​​differ greatly, the scale with the largest information gain value is selected as the representative scale of the current conditional attribute.

[0050] Step S5: Feature fusion. Select an optimal scale as the representative scale for each conditional attribute to form a feature set. Extract the value of each conditional attribute at its optimal scale to form a fusion decision table with consistent scale, which serves as a multi-scale fusion feature set.

[0051] Step S6: Establish scale dependencies. For attributes with hierarchical dependencies, model the inclusion relationship through the scale tree of the multi-scale decision table. Use depth-first traversal of the scale tree to compress redundant information and retain the most discriminative scale while ensuring decision consistency.

[0052] By performing the aforementioned operations, this solution addresses the issues of fragmented data between hospitals and other institutions, difficulties in interoperating home-based dynamic monitoring with hospital systems, and the high heterogeneity of multi-source data in terms of time granularity, sampling frequency, and semantic structure, which traditional systems lack effective fusion and dynamic modeling capabilities. Instead, it employs a multi-scale data fusion method to construct a multi-scale decision table, modeling data from different sources at different scales. Information gain is introduced as a scale selection criterion, automatically identifying the scale at which each feature has the greatest discriminative power for risk prediction, forming a high-fidelity fused feature vector. This retains the discriminative information of each data source at its optimal scale, providing a high-quality input with a clear structure and unified scale for the subsequent AI risk assessment module.

[0053] Example 4, based on the above examples, describes an AI risk assessment module that uses an improved Transformer model to capture time-series features in a multi-scale feature set. Specifically, it includes the following steps:

[0054] Step M1: Construct a temporal input sequence by organizing the multi-scale fusion features in the multi-scale fusion feature set of each pregnant woman into a temporal feature sequence in chronological order;

[0055] Step M2: Position encoding, using a cosine function to add position encoding to the temporal features in the temporal feature sequence, to obtain the position-embedded temporal feature sequence;

[0056] Step M3: Transformer encoder construction, using an L-layer Transformer encoder to process the temporal feature sequence of location embeddings. Each layer of the Transformer encoder contains a multi-head self-attention and feedforward network.

[0057] Step M4: Construct a time-aware dynamic self-attention mechanism, build a time dependency matrix and add it to the self-attention mechanism of the Transformer encoder, and calculate the time correlation using the following formula: ; ;

[0058] In the formula, and For indexing, Represents the first element in the time dependency matrix. line, number Column elements, and They represent the first The time point and the The actual timestamp of each point in time. The time decay coefficient, It is an exponential function. This is the attention weight matrix. For querying the matrix, The key matrix, The time dependency matrix, Let be the dimension of the key matrix. For normalization function, This is a transpose operation;

[0059] Step M5: Dynamic weight adjustment mechanism. After each Transformer layer, a gated dynamic weight module is introduced to adjust the weights of the Transformer encoder, resulting in an improved Transformer model.

[0060] Step M6: Rule extraction. The attention weights of the last Transformer encoder layer are used to identify the key risk time windows of the embedded temporal feature sequence, and dynamic rules are extracted by combining feature importance. The risk probability and rule interpretation are then output.

[0061] Step M7: Rule confidence and coverage evaluation. For each mined rule, calculate two metrics: rule mining accuracy and rule coverage. Optimize the improved Transformer model. Rule mining accuracy refers to the proportion of correctly predicted rules that meet the conditions; rule coverage refers to the proportion of samples that apply the rule to the total sample.

[0062] Example 5, based on the above examples, includes the following steps in step M5: The dynamic weight adjustment mechanism is further refined by the following steps:

[0063] Step M51: Global state representation construction. Calculate the global context summary for the hidden sequence output by the current layer Transformer encoder to obtain the global vector. The hidden sequence is the output of the Transformer encoder except for the last layer Transformer encoder.

[0064] Step M52: Dynamic weight vector generation, using a lightweight gated network to calculate the weight vector obtained by concatenating the hidden state vector and the global vector at the current time point;

[0065] Step M53: Hidden state update. Use the weight vector to perform a weighted fusion of the current hidden state vector and the hidden state vector at the previous time point to obtain a new hidden state vector.

[0066] Example 6, based on the above examples, in step M6, the dynamic rule takes the following form:

[0067] If, at 32 weeks of gestation, a systolic blood pressure ≥140 mmHg and a fetal heart rate variability <5 bpm for 3 consecutive days, the probability of preeclampsia is ≥0.78.

[0068] Rule 1: Gestational age ≥ 34 weeks + umbilical cord blood flow S / D > 3.5 + blood pressure ≥ 140 / 90 → Red risk (P ≥ 0.7);

[0069] Rule 2: Gestational age <34 weeks + decreased fetal movement >30% + fetal heart rate variability >25% → Orange risk (0.4≤P<0.7);

[0070] Rule 3: Gestational age ≥ 28 weeks + amniotic fluid index < 5cm + no fetal movement → emergency intervention (purple risk);

[0071] In the above rules, P is the predicted risk probability, and S / D represents the ratio of the peak systolic velocity to the end-diastolic velocity of the fetal heart.

[0072] The conditional terms in the rules are determined by high attention weights and high feature activations to ensure clinical interpretability.

[0073] Example 7, based on the above examples, describes a dynamic rule whereby the system automatically triggers a dynamic rule update when new multi-scale perinatal data arrives. The dynamic weight adjustment mechanism in step M5 is used to automatically adjust the weights of each Transformer encoder layer. The process is as follows:

[0074] Receive new multi-scale perinatal data → Calculate the matching degree between the new multi-scale perinatal data and existing rules → If the matching degree exceeds a preset threshold, trigger rule updates → Recalculate the rules using an improved Transformer model → Apply the updated rules to the system. An example update is shown below:

[0075] The system detected that among recently added cases, the probability of preeclampsia risk for those with gestational age ≥36 weeks + urine protein >300mg / 24h increased from 0.65 to 0.78. The system then automatically updated the rule: gestational age ≥36 weeks + urine protein >300mg / 24h → red risk (P≥0.75).

[0076] By performing the aforementioned operations, an improved Transformer model is designed to address the shortcomings of existing risk assessments, which are mostly based on static thresholds or fixed rules and cannot capture the dynamic evolution of physiological parameters during pregnancy, lacking individualized and adaptive modeling capabilities. This model embeds a time-dependent matrix and a dynamic weight adjustment mechanism to achieve deep modeling of time-series data and the generation of interpretable dynamic rules. It also supports online self-updating of rules, improving the accuracy and real-time performance of risk assessments, while enhancing clinical interpretability and system adaptability. The rule update cycle is shortened from several months to hours, effectively supporting individualized, dynamic, and intelligent five-color grading management of the perinatal period, providing more accurate and timely early warnings and interventions for high-risk pregnant women.

[0077] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0079] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A perinatal triage management system based on AI risk assessment, characterized in that: The system includes a data acquisition layer, a feature engineering layer, an AI risk assessment module, a five-color dynamic grading module, and an intervention push management module; The data acquisition layer connects to the hospital information system, smart wearable devices, home medical devices, mobile apps and mini-programs to collect perinatal data of pregnant women from multiple sources. The perinatal data is then layered according to different time scales to obtain multi-scale perinatal data. The feature engineering layer cleans, formats, and standardizes the collected multi-scale perinatal data, and uses a multi-scale data fusion method to fuse the multi-scale perinatal data to obtain a multi-scale feature set. The AI ​​risk assessment module uses an improved Transformer model to capture time-series features in a multi-scale feature set, generate dynamic rules, and output risk probabilities and rule interpretations. The five-color dynamic grading module automatically maps risk probabilities to five levels of risk labels: green, yellow, orange, red, and purple, and dynamically updates the risk probabilities output by the AI ​​risk assessment module. The intervention push management module establishes electronic cases for high-risk pregnant women with orange and red risk labels, automatically triggering SMS reminders, medical assistant group notifications, and referral suggestions.

2. The perinatal triage management system based on AI risk assessment according to claim 1, characterized in that: The feature engineering layer uses a multi-scale data fusion method to fuse multi-scale perinatal data. The multi-scale data fusion method specifically includes the following steps: Step S1: Decision table construction. Construct a multi-scale decision table for each type of multi-scale perinatal data. The multi-scale decision table includes an object set, a condition attribute set, and a decision attribute. All multi-scale decision tables share the same set of objects and assign a uniform decision attribute to each object. Step S2: Scale partitioning, divide each conditional attribute in the conditional attribute set into N scales according to time from fine to coarse, and construct a scale tree; Step S3: Information gain calculation, calculate the information gain of each conditional attribute in the conditional attribute set at different scales; Step S4: Scale selection. For each conditional attribute, compare the information gain at different scales. If the information gain of the current conditional attribute is similar at N scales, the coarsest scale is selected first to represent the current conditional attribute. If the information gain values ​​differ greatly, the scale with the largest information gain value is selected as the representative scale of the current conditional attribute. Step S5: Feature fusion. Select an optimal scale as the representative scale for each conditional attribute to form a feature set. Extract the value of each conditional attribute at its optimal scale to form a fusion decision table with consistent scale, which serves as a multi-scale fusion feature set. Step S6: Establish scale dependencies. For attributes with hierarchical dependencies, model the inclusion relationship through the scale tree of the multi-scale decision table. Use depth-first traversal of the scale tree to compress redundant information and retain the most discriminative scale while ensuring decision consistency.

3. The perinatal triage management system based on AI risk assessment according to claim 2, characterized in that: The AI ​​risk assessment module uses an improved Transformer model to capture time-series features from a multi-scale feature set, specifically including the following steps: Step M1: Construct a temporal input sequence by organizing the multi-scale fusion features in the multi-scale fusion feature set of each pregnant woman into a temporal feature sequence in chronological order; Step M2: Position encoding, adding position encoding to the temporal features in the temporal feature sequence to obtain the position-embedded temporal feature sequence; Step M3: Transformer encoder construction, using an L-layer Transformer encoder to process the temporal feature sequence of location embeddings. Each layer of the Transformer encoder contains a multi-head self-attention and feedforward network. Step M4: Construct a time-aware dynamic self-attention mechanism, build a time dependency matrix and add it to the self-attention mechanism of the Transformer encoder, and calculate the time correlation using the following formula: ； ； In the formula, and For indexing, Represents the first element in the time dependency matrix. line, number Column elements, and They represent the first The time point and the The actual timestamp of each point in time. The time decay coefficient, It is an exponential function. Here is the attention weight matrix. For querying the matrix, The key matrix, For time dependency matrix, Let be the dimension of the key matrix. For normalization function, This is a transpose operation; Step M5: Dynamic weight adjustment mechanism. After each Transformer layer, a gated dynamic weight module is introduced to adjust the weights of the Transformer encoder, resulting in an improved Transformer model. Step M6: Rule extraction. The attention weights of the last Transformer encoder layer are used to identify the key risk time windows of the embedded temporal feature sequence, and dynamic rules are extracted by combining feature importance. The risk probability and rule interpretation are then output. Step M7: Rule confidence and coverage evaluation. For each mined rule, calculate two metrics: rule mining accuracy and rule coverage, and optimize the improved Transformer model.

Images