CN-122025178-A - Perinatal period hierarchical management system based on AI risk assessment

Abstract

The invention relates to the technical field of medical health information, in particular to a perinatal stage classification management system based on AI risk assessment, which comprises a data acquisition layer, a characteristic engineering layer, an AI risk assessment module, a five-color dynamic classification module and an intervention pushing management module; the method comprises the steps of constructing a multi-scale decision table by adopting a multi-scale data fusion method, modeling data of different scales of different sources, introducing information gain as a scale selection standard, forming a high-fidelity fusion feature vector, retaining discriminant information of each data source under the optimal scale, providing high-quality input with clear structure and uniform scale for a subsequent AI risk assessment module, designing an improved transducer model, embedding a time-dependent matrix and a dynamic weight adjustment mechanism, realizing deep modeling of time sequence data and interpretable dynamic rule generation, supporting rule online self-updating, and improving accuracy and instantaneity of risk assessment.

Inventors

• XIONG DANDAN
• WEN QUAN
• LING YAN

Assignees

• 江西省人民医院

Dates

Publication Date

20260512

Application Date

20260202

Claims (3)

1. 1. The perinatal period grading management system based on AI risk assessment is characterized by comprising a data acquisition layer, a characteristic engineering layer, an AI risk assessment module, a five-color dynamic grading module and an intervention pushing management module; The data acquisition layer is in butt joint with a hospital information system, intelligent wearable equipment, household medical equipment, a mobile terminal APP and an applet, acquires perinatal period data of pregnant women from a multi-source channel, and layers the perinatal period data according to different time scales to obtain multi-scale perinatal period data; the characteristic engineering layer cleans, formats and standardizes the collected multi-scale perinatal data, and fuses the multi-scale perinatal data by using a multi-scale data fusion method to obtain a multi-scale characteristic set; The AI risk assessment module captures time sequence features in the multi-scale feature set by using an improved transducer model, generates dynamic rules, and outputs risk probability and rule interpretation; The five-color dynamic classification module automatically maps to green, yellow, orange, red and purple five-level risk labels according to the risk probability and dynamically updates according to the risk probability output by the AI risk assessment module; the intervention pushing management module establishes an electronic project for the high-risk pregnant women with orange and red risk labels, and automatically triggers a short message reminder, medical assistant group notification and referral advice.
2. 2. The perinatal grading management system based on AI risk assessment as set forth in claim 1, wherein the feature engineering layer uses a multi-scale data fusion method to fuse multi-scale perinatal data, the multi-scale data fusion method specifically comprises the following steps: Step S1, constructing a decision table, namely constructing a multi-scale decision table for multi-scale perinatal period data of each type, wherein the multi-scale decision table comprises an object set, a conditional attribute set and decision attributes, all the multi-scale decision tables share the same group of object set, and endowing each object with uniform decision attributes; Step S2, dividing the scale into N scales according to the time from thin to thick for each condition attribute in the condition attribute set, and constructing a scale tree; s3, calculating information gain, namely calculating the information gain of each condition attribute in the condition attribute set under different scales; Step S4, for each condition attribute, comparing information gains under different scales, if the current condition attribute is similar in information gain under N scales, preferentially selecting the coarsest scale to represent the representative scale of the current condition attribute, and if the information gain value difference is large, selecting the scale with the largest information gain value as the representative scale of the current condition attribute; Step S5, feature fusion, namely selecting an optimal scale for each condition attribute as a representative scale to form a feature set, extracting the value of each condition attribute under the optimal scale to form a fusion decision table with consistent scales, and taking the fusion decision table as a multi-scale fusion feature set; And S6, establishing a scale dependency relationship, modeling an inclusion relationship through a scale tree of a multi-scale decision table for the attribute with hierarchical dependency, traversing the scale tree by utilizing depth priority, compressing redundant information on the premise of ensuring decision consistency, and reserving the scale with the most discriminant.
3. 3. The perinatal grading management system based on AI risk assessment as set forth in claim 2 wherein the AI risk assessment module captures time series features in a multi-scale feature set using a modified transducer model, specifically comprising the steps of: M1, constructing a time sequence input sequence, and organizing the multi-scale fusion features in the multi-scale fusion feature set of each pregnant woman into a time sequence feature sequence according to time sequence; M2, position coding, namely adding position coding to the time sequence features in the time sequence feature sequence to obtain a position embedded time sequence feature sequence; step M3, constructing a transducer encoder, namely embedding a time sequence characteristic sequence by using processing positions of an L-layer transducer encoder, wherein each layer of the transducer encoder comprises a multi-head self-attention and feedforward network; And M4, constructing a time-aware dynamic self-attention mechanism, constructing a time-dependent matrix, adding the time-dependent matrix into the self-attention mechanism of the transducer encoder, and calculating time correlation by using the following formula: ; ; In the formula, And For the purpose of the index, Representing the th in a time dependent matrix Line 1 The elements of the column are arranged such that, And Respectively represent the first Time point and the first The actual time stamp of the point in time, As a coefficient of the decay in time, As a function of the index of the values, In order to pay attention to the weight matrix, In order to query the matrix, In the form of a matrix of keys, In the form of a time-dependent matrix, For the dimensions of the key matrix, As a function of the normalization, Is a transposition operation; Step 5, a dynamic weight adjustment mechanism, after each transducer layer, introducing a gate control dynamic weight module to adjust the weight of a transducer encoder so as to obtain an improved transducer model; M6, extracting rules, namely embedding a key risk time window of a time sequence feature sequence into the attention weight identification position of the last layer of transducer encoder, extracting dynamic rules by combining feature importance, and outputting risk probability and rule interpretation; and M7, evaluating the confidence coefficient and the coverage degree of the rule, calculating two indexes of rule mining accuracy and rule coverage rate for each mined rule, and optimizing the improved transducer model.

Description

Perinatal period hierarchical management system based on AI risk assessment Technical Field The invention relates to the technical field of medical health information, in particular to a perinatal stage classification management system based on AI risk assessment, which is suitable for full-period high-risk screening, dynamic early warning and intervention support of pregnant and lying-in women from pre-pregnant to post-parturient 42 days. Background Perinatal management refers to systematic health monitoring and intervention of pregnant and parturients and newborns in the key stage of 28 weeks to 1 week after pregnancy, and has the core aims of early identification of high-risk factors, prevention and reduction of complications and death of the pregnant and parturients and newborns, guarantee of maternal and infant safety, and improvement of birth population quality, and is an important component of a maternal and child health care system. At present, the perinatal management of China still faces multiple challenges that the identification of high-risk pregnant and lying-in women is highly dependent on subjective experience of clinicians, different doctors have obvious differences in interpretation of the same indexes (such as blood pressure, fetal heart variation and ultrasonic parameters), meanwhile, data in a hospital are severely fractured from data outside the hospital, dynamic health data (such as remote fetal heart monitoring, blood pressure, blood sugar and fetal movement records) collected at home are difficult to realize real-time intercommunication and fusion with information systems such as hospitals HIS, LIS, PACS, more importantly, various data are high in isomerism on time granularity, sampling frequency and semantic structure and difficult to effectively integrate in a traditional system, and in addition, the existing risk assessment is mainly based on static threshold values or fixed rules (such as 'blood pressure is more than or equal to 140/90 is high risk'), so that dynamic evolution of physiological parameters in a pregnancy process cannot be captured, and accurate and real-time risk early warning and grading intervention are difficult to realize. Disclosure of Invention Aiming at the problems that in the prior art, home dynamic monitoring is difficult to communicate with a hospital system and multisource data is highly heterogeneous in time granularity, sampling frequency and semantic structure, and the traditional system lacks effective fusion and dynamic modeling capacity, the scheme adopts a multiscale data fusion method to construct a multiscale decision table, models data of different scales of different sources, introduces information gain as a scale selection standard, automatically identifies which scale each feature has most discrimination on risk prediction, forms a high-fidelity fusion feature vector, reserves discriminant information of each data source under the optimal scale, provides high-quality input with clear structure and uniform scale for a subsequent AI risk assessment module, and aims at the problems that the traditional risk assessment is based on a static threshold or a fixed rule, cannot capture dynamic evolution of physiological parameters in a pregnancy process, lacks individuation and self-adaption modeling capacity, designs an improved transformor model, embeds a time-dependent matrix and dynamic weight adjustment, realizes depth and scale interpretation and high-precision updating and support of the dynamic modeling rule, and realizes accurate and real-time sequence interpretation and online generation of the dynamic rule. The invention provides a perinatal stage classification management system based on AI risk assessment, which comprises a data acquisition layer, a characteristic engineering layer, an AI risk assessment module, a five-color dynamic classification module and an intervention pushing management module, wherein the data acquisition layer is used for acquiring the AI risk of the perinatal stage classification management system; The data acquisition layer is in butt joint with a hospital information system, intelligent wearable equipment, household medical equipment, a mobile terminal APP and an applet, acquires perinatal period data of pregnant women from a multi-source channel, and layers the perinatal period data according to different time scales to obtain multi-scale perinatal period data; the characteristic engineering layer cleans, formats and standardizes the collected multi-scale perinatal data, and fuses the multi-scale perinatal data by using a multi-scale data fusion method to obtain a multi-scale characteristic set; The AI risk assessment module captures time sequence features in the multi-scale feature set by using an improved transducer model, generates dynamic rules, and outputs risk probability and rule interpretation; The five-color dynamic classification module automatically maps to green, yellow, orange, red and purple fi