Perinatal whole-cycle intelligent intervention method based on multi-source heterogeneous data fusion

Abstract

This invention provides a perinatal full-cycle intelligent intervention method based on multi-source heterogeneous data fusion, comprising: step S100, simultaneously collecting subjective self-reported data from pregnant women, real-time physiological data from wearable devices, and heterogeneous clinical examination data to construct a multi-dimensional spatiotemporal data pool; step S200, establishing a unified time base axis with gestational age and absolute time as dual coordinate axes, and extracting feature vectors from the data in the multi-dimensional spatiotemporal data pool at fixed time steps; step S300, calculating the real-time global comprehensive risk index RI based on the feature vectors; step S400, when RI exceeds the first dynamic threshold T1, the system triggers a level-one warning; when RI exceeds the second dynamic threshold T2 and a preset acute-critical feature operator is identified, the system triggers a level-two warning, and simultaneously generates an admission preparation list and sends a rescue instruction to the hospital's emergency dispatch system.

Description

Technical Field

[0001] This invention relates to a medical data management method, and more particularly to an intelligent intervention method for the entire perinatal period based on the fusion of multi-source heterogeneous data. Background Technology

[0002] With the widespread adoption of IoT devices and mobile healthcare, perinatal health monitoring for pregnant women is gradually expanding from simple in-hospital clinical examinations to routine out-of-hospital monitoring. Current perinatal health monitoring systems largely rely on pregnant women's home physiological devices (such as blood pressure monitors and blood glucose meters) and mobile app records. However, existing perinatal monitoring and intervention technologies still have the following significant defects in practical applications: (1) Existing technologies usually treat the subjective self-reported data of pregnant women, the high-frequency streaming data of wearable devices and the discrete examination data of hospitals separately. Since these multi-source heterogeneous data have huge differences in collection frequency and time attributes, most existing systems can only make isolated single-point judgments; (2) When the current health risk assessment model triggers an alarm, it can often only output a vague "high risk" signal, and cannot accurately trace and clearly point out the core pathogenic or risk factors that caused the warning. This makes doctors still need to spend a lot of time to review the original data to investigate the cause after receiving the warning, which reduces the timeliness of clinical intervention; (3) Most existing intelligent monitoring systems only push warning information to users or doctors. The system functions are limited to simple "health care" and monitoring. When encountering acute and critical illnesses such as antepartum hemorrhage and severe preeclampsia, the system cannot predict and dispatch the hospital's emergency channels, rescue consumables and blood bank resources in advance according to the risk level, and cannot achieve a true closed loop from information perception to clinical process control. Summary of the Invention

[0003] To address the aforementioned problems, this invention provides a method for intelligent intervention throughout the perinatal period based on multi-source heterogeneous data fusion, comprising:

[0004] Step S100: Simultaneously collect subjective self-reported data from pregnant women, real-time physiological data from wearable devices, and heterogeneous data from clinical examinations to construct a multi-dimensional spatiotemporal data pool;

[0005] Step S200: Establish a unified time base axis with gestational age and absolute time as dual coordinate axes, and extract feature vectors from the data in the full-dimensional spatiotemporal data pool according to a fixed time step.

[0006] Step S300: Calculate the real-time global comprehensive risk index RI based on the feature vector;

[0007] In step S400, when the RI exceeds the first dynamic threshold T1, the system triggers a level one warning. When the RI exceeds the second dynamic threshold T2 and a preset critical feature operator is identified, the system triggers a level two warning. At the same time as generating the admission preparation list, the system sends a rescue instruction to the hospital's emergency dispatch system.

[0008] Furthermore, the specific process of step S200 includes:

[0009] Step S210: Establish a unified time base axis using gestational age and absolute time as dual coordinate axes, and mark the time on the unified time base axis for the collected heterogeneous data.

[0010] Step S220: Downsample the real-time physiological data from the wearable device and perform forward padding on the subjective self-reported data and hospital interface data;

[0011] Step S230: Create a feature vector containing three types of attributes within each time step: V t =[S self (t),W wear (t),C clinical (last_val)].

[0012] Furthermore, the process of downsampling the real-time physiological data of the wearable device in step S220 includes:

[0013] Step S221: Remove all frequency components higher than half of the new sampling rate using a low-pass filter;

[0014] Step S222: Take values ​​according to the set step size M, and retain the maximum, minimum and average values ​​within the step size.

[0015] Furthermore, the forward-filling process for the subjective self-reported data and hospital interface data in step S220 includes:

[0016] Step S223: Determine the impact window for each data point;

[0017] Step S224, set the step size M, with the initial step size M start Obtain subjective self-reported data and hospital interface data in subsequent time segments {M start+1 M start+2 The function checks if the data field in the sequence M is empty. If it is empty and within the validity period, the initial time segment M is set to M. start The value is copied to the data field of the current time slice; if it is not empty, the time slice is set as the initial time slice T. start ;

[0018] Step S225: Calculate the confidence coefficient C(t) for the data in each time segment. If C(t) is less than the threshold, remeasure the subjective self-reported data and the hospital interface data.

[0019] ,

[0020] Where t is the time difference since the last measurement, C0 is the original confidence level, and T valid λ represents the clinical efficacy period, λ is the decay factor, and e is the natural logarithm.

[0021] Furthermore, the specific process of step S300 includes:

[0022] Step S310: Decouple the feature vector into a clinical examination feature subspace, a wearable device physiological feature subspace, and a subjective self-report feature subspace;

[0023] Step S320: Construct a clinically guided cross-attention mechanism network, and use linear mapping to integrate clinical examination features F. clinical Transform into query vector matrix Q c The physiological characteristics F of the wearable device wear Compared with subjective self-reported characteristics F self Transform them into the corresponding key vector matrix K respectively. w K s Sum value vector matrix V w V s ;

[0024] Step S330: Calculate the scaled dot product of the query vector matrix and the two key vector matrices respectively, and obtain the cross-attention weight matrix after normalization by the Softmax function.

[0025] Step S340: The corresponding value vector matrix is ​​weighted and summed using the cross-attention weight matrix to obtain physiological context features and behavioral context features. The clinical examination features, physiological context features and behavioral context features are then concatenated into a multi-dimensional feature tensor to generate a global fusion feature tensor.

[0026] Step S350: Input the global fusion feature tensor into the multilayer perceptron network for nonlinear mapping and output the global comprehensive risk index RI in real time.

[0027] Furthermore, in step S330, Q c With K w The scaled dot product, and Q c With K s The dot product is shown in the following formula.

[0028] ,

[0029] ,

[0030] Where, d k is the dimension of the key vector.

[0031] Furthermore, step S350 involves obtaining the global comprehensive risk index RI through nonlinear mapping using a multilayer perceptron network.

[0032] ,

[0033] Among them, F fusion For the global fusion feature tensor, For nonlinear mapping, W out and b out σ represents the output layer weights and biases learned by the model during the pre-training phase, and σ is the activation function.

[0034] Furthermore, the output layer weights W out and bias b out The learning process is as follows:

[0035] Step S351: Extract features from the subjective self-reported data, real-time physiological data and clinical examination heterogeneous data of the input sample X to form a feature vector, and label the pregnancy results in the input sample with a label Y, with adverse results recorded as 1 and otherwise as 0;

[0036] Step S352, randomly generate output layer weights W out and bias b out ;

[0037] Step S353: Input feature vector, through attention mechanism and multilayer perceptron calculation, by W out and b out The mapping yields a predicted risk RI predict ;

[0038] Step S354: Use cross-entropy loss to calculate the gap between the predicted risk and the true outcome label Y;

[0039] Step S355: Using the chain rule, calculate the loss function for W. out and b out The gradient;

[0040] Step S356: Using the Adam optimizer, update W in the opposite direction of the gradient. out and b out ,

[0041] ,

[0042] Loss is the cross-entropy loss function, and η is the learning rate;

[0043] Step S357: Repeat steps S353 to S356 until the loss function no longer decreases.

[0044] Furthermore, in step S400, the secondary early warning obtains rescue instructions through a resource prediction algorithm, which specifically includes:

[0045] Step S410: Receive the global comprehensive risk index RI and the risk contribution factor identifier F. id And the identified critical characteristic operators, risk contribution factor identifier F id The cross-attention weight matrix includes critical feature operators such as fetal distress operator, severe preeclampsia operator, and antepartum hemorrhage risk operator.

[0046] Step S420: Construct the correspondence between risk contribution factor identifiers of different dimensions and medical resource demand types, wherein the medical resource demand types include at least bed resources B, blood resources L, emergency green channel resources P, and emergency consumable resources E;

[0047] Step S430: Call the preset resource prediction function RP=f(RI,F id G week ), calculate the probability and quantity of demand for various types of medical resources, G week This is gestational age data;

[0048] Step S440: Based on the prediction results generated by the resource prediction algorithm, the pre-occupancy logic marking of medical resources is executed through the standard interface of the Hospital Management System (HIS). The pre-occupancy logic marking includes: generating a reserved bed identifier and locking the target bed in the bed management system; sending a pre-request instruction for a specific blood type in the blood management system; automatically generating an admission preparation list that matches the risk contribution factor and pushing it to the medical staff terminal.

[0049] Compared with the prior art, the present invention has the following advantages: (1) It establishes a dual coordinate axis with gestational age and absolute time, which solves the problem of alignment of data in different dimensions in terms of developmental logic and life rhythm; at the same time, it constructs a full-dimensional spatiotemporal data pool by downsampling high-frequency wearable data and performing forward filling based on the influence effect window for discrete clinical data; (2) It uses the cross-attention mechanism of clinical guidance to actively screen abnormal segments in wearable and self-reported data and form risk contribution factor identifiers by using clinical examination features as queries. While outputting the risk index RI, it will clearly indicate which dimension combinations caused the current risk; (3) When the preset acute and critical feature operator is identified, the system will no longer just alarm, but directly trigger a secondary warning. The system automatically generates an admission preparation list and sends a rescue instruction to the emergency dispatch system through API. At the same time, it executes the "pre-occupancy logic mark" of medical resources (beds, blood, etc.).

[0050] The present invention will now be further described with reference to the accompanying drawings. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0052] Figure 2 This is a schematic diagram of a dynamic risk assessment model. Detailed Implementation

[0053] Combination Figure 1 The system upon which this method is based includes a feature decoupling module, a guided cross-attention mechanism module, a feature fusion module, and an output module. The feature decoupling module receives subjective self-reported data from pregnant women and real-time physiological data from wearable devices, maps the heterogeneous clinical examination data in real-time via a standard hospital interface, and extracts features to construct feature vectors. The guided cross-attention mechanism module incorporates a dynamic risk assessment model. This model uses a cross-attention mechanism network to guide the application of heterogeneous clinical examination data to subjective self-reported data and real-time physiological data, identifying which dynamic features have higher risk contribution weights in the current clinical context. The feature fusion module concatenates the features that have undergone cross-attention (symbols in the figure). The output module performs nonlinear mapping and dimensionality reduction on the spliced ​​tensor input multilayer perceptron (MLP) to obtain the quantized global comprehensive risk index RI.

[0054] The subjective self-reported data in this embodiment includes fetal movement self-monitoring, symptom self-reporting, lifestyle records, and medication adherence feedback. Fetal movement self-monitoring includes the number of fetal movements per hour and the strength of those movements. Symptom self-reporting includes the nature of abdominal pain (dull / painful), vaginal bleeding (present / absent, magnitude), dizziness, blurred vision, and the degree of limb edema. Lifestyle records include immediate dietary intake (carbohydrate / sugar description), emotional state score (anxiety / depression perception), and immediate physical activity description. Medication adherence feedback includes insulin injection records and antihypertensive medication usage. Real-time physiological data includes fetal dimensions, maternal baseline vital signs, exercise and metabolic indicators, and waveform characteristics. Fetal dimensions include real-time fetal heart rate (FHR) baseline, acceleration / deceleration frequency, and fetal heart rate variability. Maternal baseline vital signs include real-time heart rate (HR), heart rate variability (HRV), blood oxygen saturation (SpO2), and ambulatory blood pressure (systolic / diastolic). Exercise and metabolic indicators include real-time steps, calories burned, resting heart rate, and sleep quality score (deep sleep / light sleep duration). Heterogeneous clinical examination data includes basic medical history and physical signs, biochemical laboratory indicators, imaging and functional examination results, and clinical diagnostic operators. Basic medical history and physical signs include gestational age, expected delivery date, height, pre-pregnancy weight, obstetric history (G / P), and past medical conditions (such as hypertension and diabetes). Biochemical laboratory indicators include fasting blood glucose, oral glucose tolerance test (OGTT) results, hemoglobin (HGB), coagulation function indicators (PT / APTT), urinary protein quantification, and serum ferritin. Imaging and functional examination results include ultrasound measurements (biparietal diameter, abdominal circumference, femur length), amniotic fluid index (AFI), umbilical artery flow ratio (S / D), placental grade and location (e.g., risk of placenta previa). Clinical diagnostic operators include gestational diabetes mellitus (GDM) labels, gestational hypertension (PIH) classifications, and risk indicators for scarred uterus. The above data can reflect abnormal conditions during the perinatal period. Abnormal conditions and warning logic are explained below, but are not limited to the situations listed in this embodiment.

[0055] (1) Fetal distress: Fetal heart rate (FHR) baseline, acceleration / deceleration frequency, fetal heart rate variability, number of fetal movements, and strength of fetal movements reflect the fetal distress situation. When the fetal heart rate baseline is higher than or lower than the threshold, or when there are N late decelerations within a certain period of time, accompanied by significantly less fetal movement than the threshold, it indicates fetal hypoxia.

[0056] (2) Fetal growth restriction: Ultrasound measurements and umbilical artery blood flow ratio (S / D) reflect the degree of fetal growth restriction; when the umbilical artery blood flow ratio continues to rise within a certain period of time, it indicates that the placental peripheral resistance is high and insufficient blood supply leads to fetal growth retardation.

[0057] (3) Amniotic fluid abnormalities: Amniotic fluid index (AFI) and maternal blood oxygen saturation (SpO2) reflect the abnormality of amniotic fluid; when the amniotic fluid index is less than 5cm, it indicates that the amniotic fluid is too low, which can easily lead to umbilical cord compression and cause real-time fetal heart rate fluctuations.

[0058] (4) Preeclampsia / preeclampsia: Ambulatory blood pressure (systolic / diastolic), quantitative urine protein, and limb edema reflect preeclampsia / preeclampsia; when blood pressure is consistently higher than 140 / 90 mmHg for a certain period of time and urine protein is positive, it indicates preeclampsia.

[0059] (5) HELLP syndrome: Clinical diagnostic operators, coagulation function indicators (PT / APTT), and hemoglobin (HGB) reflect HELLP syndrome; in the context of hypertension, if coagulation function indicators and hemoglobin are less than the threshold within a certain period of time, it indicates elevated blood and liver enzymes.

[0060] (6) Hyperglycemia: The results of the oral glucose tolerance test (OGTT), immediate dietary intake (carbohydrate / sugar), and insulin execution feedback reflect hyperglycemia; if the results of the oral glucose tolerance test (OGTT) exceed the threshold, or if the insulin dose and effect are inversely proportional, it indicates blood glucose fluctuation.

[0061] (7) Risk of placenta previa / placental abruption: Placental location / grading, vaginal bleeding (level), and nature of abdominal pain (pain) can reflect the possible risk of placenta previa / placental abruption; if imaging suggests low-lying placenta, or if the patient subjectively reports "painless bleeding", then the risk of placenta previa is high.

[0062] (8) Risk of uterine rupture: Risk markers for scarred uterus, pregnancy history (G / P), and the nature of abdominal pain can reflect the risk of uterine rupture; if a pregnant woman with a scarred uterus experiences abdominal pain for a period of time, the thickness of the lower uterine segment and changes in fetal heart rate should be closely monitored.

[0063] Combination Figure 1 A perinatal full-cycle intelligent intervention method based on multi-source heterogeneous data fusion, implemented according to the above system, includes the following steps:

[0064] Step S100: Collect the pregnant woman’s subjective self-reported data and the wearable device’s real-time physiological data simultaneously through the user terminal, and map the heterogeneous clinical examination data in real time through the hospital’s standard interface to build a full-dimensional spatiotemporal data pool.

[0065] Step S200: The unstructured information in the data pool is transformed into a structured data frame with temporal alignment features using the normalization unit, and a feature vector containing the joint weights of physiological evolution trends, behavioral deviations and clinical indicators is extracted through the multimodal feature extraction unit.

[0066] Step S300: Input the feature vector into a pre-trained dynamic risk assessment model. The model identifies risk contribution factors in each dimension based on an attention mechanism and outputs the global comprehensive risk index RI in real time.

[0067] In step S400, when the RI exceeds the first dynamic threshold T1, the system triggers a level one warning. When the RI exceeds the second dynamic threshold T2 and a preset critical characteristic operator is identified, the system triggers a level two linkage mechanism. While generating the admission preparation list, the system sends a rescue instruction to the hospital emergency dispatch system through the API interface. The level two linkage mechanism includes a resource prediction algorithm that predicts the possible bed and blood bank demand based on the current risk assessment results and pre-occupies the space in the hospital resource management system.

[0068] Step S500: Collect pregnancy outcome data of the institution in real time as feedback labels, and use incremental learning algorithm to perform localized fitting of model parameters and threshold sequence {T1,T2} to achieve self-evolution of evaluation model.

[0069] In step S100, subjective self-reported data, wearable device real-time physiological data, and hospital interface data constitute heterogeneous data. Subjective self-reported data is low-frequency, unstructured data; wearable device real-time physiological data is high-frequency streaming data; and hospital interface data is discrete data. Therefore, it is necessary to address the access standards for these three data streams, specifically including:

[0070] (1) Subjective self-reported data such as fetal movement count, emotional state, and abdominal pain were input through the scale or voice input of the mobile APP. Natural language processing (NLP) was used to convert the unstructured description into numerical degree labels.

[0071] (2) Real-time physiological data such as heart rate, blood oxygen, and sleep are transmitted back to wearable devices via Bluetooth or Wi-Fi. The sliding window algorithm is used to perform feature dimensionality reduction on the original waveform and extract statistical features such as mean and variance.

[0072] (3) Connect to the hospital's HIS / LIS / EMR system via HL7 or FHIR protocol to obtain hospital interface data such as blood test and ultrasound imaging results.

[0073] In step S200, since the wearable device transmits high-frequency data per second, while hospital examinations may be conducted every few weeks and self-reported data is irregular, it is necessary to align the heterogeneous data. The specific process includes:

[0074] Step S210: Establish a unified time base axis using gestational age and absolute time as dual coordinate axes, and mark the time on the collected heterogeneous data on the unified time base axis.

[0075] Step S220: Downsample the real-time physiological data from the wearable device and perform forward padding on the subjective self-reported data and hospital interface data;

[0076] Step S230: Create a feature vector containing three types of attributes within each time step: V t =[S self (t),W wear (t),C clinical (last_val)].

[0077] In step S210, absolute time alone cannot describe the developmental logic of the fetus, while relative time such as gestational age alone cannot capture the rhythm of a pregnant woman's daily life. The method of establishing a unified time base axis using gestational age and absolute time as dual coordinate axes involves assigning two time labels to each data point. When reading wearable data, the system first determines whether the pregnant woman is in a resting state using absolute time. After determining the state, it queries the corresponding gestational age. If the gestational age is 32 weeks, the system will automatically retrieve the physiological norms for 32 weeks. For example, if the pregnant woman's blood pressure reading is the same as last week (i.e., the absolute value has not changed), but relative to her gestational age of one week (i.e., the relative time has increased), this cessation of growth may actually indicate limited placental function.

[0078] Because wearable devices generate raw physiological data at extremely high frequencies, such as ECG sampling rates typically above 250Hz, directly storing real-time physiological data from wearable devices into the spatiotemporal data pool would lead to storage redundancy and computational overload. Therefore, step S220 involves downsampling the real-time physiological data from wearable devices before taking values. The specific process includes:

[0079] Step S221: Remove all frequency components higher than half of the new sampling rate using a low-pass filter;

[0080] Step S222: Take values ​​according to the set step size M, and retain the maximum, minimum and average values ​​within the step size.

[0081] Wearable devices typically use very high raw sampling rates to capture fine waveforms; however, for perinatal risk assessment, it's not necessary to look at voltage changes every millisecond, but rather to focus on minute-level heart rate trends; therefore, downsampling is needed to reduce the data from high frequency. In step S221, the new sampling rate f... new With the sampling frequency f of wearable devices wearables The relationship is f new =f wearables / M, where M is the extraction step size.

[0082] In step S222, when the signal fluctuates violently, such as when a suspected uterine contraction waveform is detected, the M value is automatically reduced to improve the sampling resolution; when the signal is in a stable period, such as when the pregnant woman is in deep sleep, the M value is increased to reduce the sampling resolution to save system resources. Retaining the maximum, minimum, and average values ​​within this step size is crucial for capturing transient tachycardia in pregnant women. Transient tachycardia, also known as palpitations, is characterized by its short duration and strong bursts. If only the average value is retained, the palpitation signal may be smoothed out. Retaining the maximum value, even if the average value is normal, will trigger a risk warning for arrhythmia or transient palpitations if the system detects a maximum value greater than the threshold.

[0083] The biggest challenge in processing subjective self-reported data such as fetal movement counts, emotional state, and abdominal pain sensations, as well as hospital-interface data such as urine protein, hemoglobin, and glucose tolerance, lies in their high dispersion; that is, a single blood test result may need to support risk assessments for several weeks. To integrate these discrete hospital-interface data into a spatiotemporal data pool, the following steps are required to achieve influence mapping:

[0084] Step S223: Determine the impact window of each data point. The impact window refers to the lifecycle of the data's backward mapping, i.e., the validity of the data for subsequent judgments. For example, the blood pressure value from 3 weeks ago is not very meaningful for the current risk assessment, but the blood type data from 3 weeks ago is still valid.

[0085] Step S224, set the step size M, with the initial step size M start Obtain subjective self-reported data and hospital interface data in subsequent time segments {M start+1 M start+2 The function checks if the data field in the sequence M is empty. If it is empty and within the validity period, the initial time segment M is set to M. start The value is copied to the data field of the current time slice; if it is not empty, the time slice is set as the initial time slice T. start ;

[0086] Step S225: Calculate the confidence coefficient C(t) for the data in each time segment, where t is the time difference from the last measurement. If C(t) is less than the threshold, remeasure the subjective self-reported data and the hospital interface data.

[0087] ,

[0088] Where C0 is the original confidence level, usually 1.0; T validThe clinical validity period is given by λ, which is the decay factor obtained through a pre-defined confidence coefficient and the clinical validity period; e is the natural logarithm. λ is not a fixed hyperparameter, but a dynamic adaptive parameter based on clinical dynamics; for example, oral glucose tolerance test (OGTT) results are slow variables, and their corresponding decay factor λ... OGTT The effect is minimal, declining only slowly over several months; while hemoglobin or coagulation function are moderate variables, declining slightly faster.

[0089] The dynamic risk assessment model in step S300 includes a feature decoupling module ( Figure 2 Feature Decoupling module), Guided cross-attention module ( Figure 2 Guided Cross-Attention module and feature fusion module ( Figure 2 Feature Fusion module), output module ( Figure 2 The Output module (in the middle) uses a linear projection layer to decouple feature vectors in a structured data frame with temporally aligned features into three independent feature subspaces. The guided cross-attention module includes two cross-attention mechanism modules ( Figure 2 The Cross-Attention module uses clinical examination features as queries to guide the selection of physiological features from wearable devices and subjective self-reported features. The feature fusion module concatenates the outputs of the Cross-Attention module to obtain a global fusion feature tensor. The output module processes the global fusion feature tensor using a multilayer perceptron to obtain the global comprehensive risk index RI. The guided Cross-Attention mechanism transforms discrete clinical examination features into query vectors to actively query and filter abnormal segments in wearable device and self-reported data. For example, in a clinical context of gestational hypertension, the model automatically assigns higher attention weights to subtle blood pressure fluctuations or heart rate variability captured by the wearable device, thus avoiding potential risks that might be overlooked in a normal context. Simultaneously, through the calculated attention weight matrix, the system, while outputting the risk index RI, can clearly indicate which combination of dimensions caused the current risk.

[0090] Based on the dynamic risk assessment model described above, step S300 specifically includes the following processes:

[0091] Step S310: Decouple the feature vectors in the structured data frame with temporal alignment features into three independent feature subspaces: clinical examination feature subspace F clinical Wearable device physiological characteristic subspace F wear and the subjective self-reported feature subspace F self ;

[0092] Step S320: Construct a clinically guided cross-attention mechanism network, and use linear mapping to transfer the clinical examination features F clinical Transform into a query vector matrix Query(Q) c ), and the physiological characteristics F of the wearable device wear Compared with subjective self-reported characteristics F self Transform them into the corresponding key vector matrix Key(K) respectively. w K s ) and value vector matrix Value(V) w V s );

[0093] Step S330: Calculate the scaled dot product of the query vector matrix and the two key vector matrices respectively, and obtain the cross-attention weight matrix after normalization using the Softmax function; the cross-attention weight matrix represents the risk contribution factor F corresponding to real-time physiological fluctuations and subjective behavioral deviations under specific clinical baseline data. id ;

[0094] Step S340: The corresponding value vector matrix is ​​weighted and summed using the cross-attention weight matrix to obtain physiological context features and behavioral context features. These features are then combined with the clinical examination feature subset to generate a global fusion feature tensor.

[0095] Step S350: Input the global fusion feature tensor into the multilayer perceptron network for nonlinear mapping and output the global comprehensive risk index RI in real time.

[0096] Step S330 specifically includes the following process:

[0097] Step S331: Obtain the clinical-physiological cross-attention matrix using a scaled dot product model, specifically by calculating Q. c With K w The dot product is used to assess which physiological fluctuations (such as palpitations) the model should assign higher attention scores to, given the current specific clinical history or biochemical indicators.

[0098] ,

[0099] Where, d k The dimension of the key vector;

[0100] Step S332: Obtain the clinical-self-reported cross-attention matrix using a scaled dot product model, specifically by calculating Q. c With K s The dot product is used to assess the potential risk weights of subjective perceptions (such as abdominal pain and fetal movement) in the current clinical context.

[0101] ,

[0102] The calculated attention weight matrix constitutes the risk contribution factor identifier in the model output;

[0103] Step S333: The weighted context vector Attention wear Attention self Compared with the original clinical features F clinical Residual connections and multidimensional splicing are performed to form a global fusion feature tensor.

[0104] The global fusion feature tensor is input into a multilayer perceptron (MLP) for nonlinear dimensionality reduction mapping, and a global comprehensive risk index RI between 0 and 1 (or a specific range) is output in real time through an activation function.

[0105] In step S331, the dot product calculates the similarity or correlation between two vectors in a multidimensional space. Clinical data Q c With a specific pathological background (such as "gestational hypertension"), it interacts with countless physiological fragments collected around the clock by wearable devices. w By comparing them one by one, if a certain physiological segment (such as a small fluctuation in blood pressure or a heart rate variation over a few minutes) closely matches the pathological characteristics of "gestational hypertension" in the feature space, their dot product will be very large. In other words, small fluctuations in blood pressure in the context of hypertension need to be taken seriously.

[0106] In steps S331 and S332, the dot product result is scaled (divided by). After processing with the softmax function, an attention weight matrix is ​​generated. The values ​​in this matrix are between 0 and 1, which represent the guidance weights that clinical data assign to wearable / self-reported data.

[0107] In step S350, the final acquisition of RI is achieved by a multilayer perceptron (MLP) through a specific activation function, the logical formula of which is:

[0108]

[0109] F fusion For global fusion feature tensor; This is a non-linear mapping, representing the processing procedure of the hidden layer in an MLP; W out and b out σ represents the output layer weights and biases learned by the model during the pre-training phase; σ is the activation function, typically the Sigmoid function.

[0110] In step S350, the output layer weights W out and bias b out The learning process is as follows:

[0111] Step S351: Extract features from the subjective self-reported data, real-time physiological data and clinical examination heterogeneous data of the input sample X to form a feature vector, and label the pregnancy results in the input sample with a label Y. If adverse results such as eclampsia, fetal distress, or premature birth occur, record it as 1; otherwise, record it as 0.

[0112] Step S352, randomly generate output layer weights W out and bias b out ;

[0113] Step S353: Input feature vector, process through attention mechanism and multilayer perceptron (MLP) calculation, and finally W... out and b out The mapping yields a predicted risk RI predict ;

[0114] Step S354: Use cross-entropy loss to calculate the gap between the predicted risk and the true outcome label Y;

[0115] Step S355: Using the chain rule, calculate the loss function for W. out and b out The partial derivative (i.e., gradient) represents the direction of parameter adjustment in order to reduce the error;

[0116] Step S356: Using the Adam optimizer, update W in the opposite direction of the gradient. out and b out ,

[0117] ,

[0118] Loss is the cross-entropy loss function, and η is the learning rate;

[0119] Step S357: Repeat steps S353 to S356 until the loss function no longer decreases.

[0120] The critical feature operators in step S400 refer to specific pattern combinations extracted from multimodal feature vectors that can characterize perinatal emergencies. In this embodiment, the critical feature operators include fetal distress operators, severe preeclampsia operators, and antepartum hemorrhage risk operators. The identification data corresponding to each critical feature operator is shown below:

[0121] (1) Fetal distress operator: Fetal movement decreased or disappeared during the cycle in subjective self-reported data, and the baseline fetal heart rate in real-time physiological data exceeded the normal threshold or there was periodic late deceleration;

[0122] (2) Severe preeclampsia operator: strong positive urine protein or blood biochemical indicators exceeding the normal threshold in clinical examination data, and blood pressure continuously rising in a short period of time in real-time physiological data;

[0123] (3) Risk factor for antepartum hemorrhage: placenta previa in clinical examination data, painless vaginal bleeding in subjective self-reported data, and increased rate and changes in blood oxygen value in real-time physiological data.

[0124] Only when the operator is hit will the system determine that immediate rescue is needed, thereby triggering a two-level linkage, automatically generating an admission preparation list, and sending rescue instructions to the hospital dispatch system.

[0125] The resource prediction algorithm in step S400 specifically includes:

[0126] Step S410: Receive the global comprehensive risk index RI and risk contribution factor identifier F generated in real time by the dynamic risk assessment model. id And the identified critical feature operators;

[0127] Step S420: Construct the resource mapping matrix M res Resource mapping matrix M res This is used to pre-determine the correspondence between risk contribution factors of different dimensions and types of medical resource demand; among which, the types of medical resource demand include at least bed resources (B), blood resources (L), emergency green channel resources (P), and emergency consumable resources (E).

[0128] Step S430: Call the preset resource prediction function RP=f(RI,F id G week ), calculate the probability and quantity of demand for various types of medical resources, G week This is gestational age data;

[0129] Step S440: Based on the prediction results generated by the resource prediction algorithm, the pre-occupancy logic marking of medical resources is executed through the standard interface of the Hospital Management System (HIS). The pre-occupancy logic marking includes: generating a reserved bed identifier and locking the target bed in the bed management system; sending a pre-request instruction for a specific blood type in the blood management system; automatically generating an admission preparation list that matches the risk contribution factor and pushing it to the medical staff terminal.

[0130] In step S420, the risk contribution factor identifier F id This is the output of the cross-attention mechanism module, which identifies the main factors leading to the current risk, based on the received risk contribution factor identifier F. id In the resource mapping matrix M res In a targeted search, find results related to F. id The corresponding specific resource dimension.

[0131] The specific process of step S430 includes:

[0132] Step S431, based on the risk contribution factor identifier F id The corresponding resource dimension is retrieved from the resource mapping matrix; for example, when the risk contribution factor points to the risk of antepartum hemorrhage, the calculation weights of blood resources L and emergency consumables E are activated.

[0133] Step S432: Combine the global comprehensive risk index RI with the pregnant woman's current gestational age data G. week Nonlinear coupling is used to determine the urgency of resource demand and baseline holdings;

[0134] Step S433: Calculate the expected consumption of target resources using the cluster center coordinates of the critical feature operators; for example, by identifying the severe eclampsia operator, match the corresponding standardized blood and medication pathways for emergency treatment, and output the quantified blood preparation units and bed levels.

[0135] Step S500 specifically includes the following processes:

[0136] Step S510: Trace the final pregnancy outcome of the monitored pregnant woman and convert the pregnancy outcome into a binary label Y. If adverse outcomes such as eclampsia, fetal distress, premature birth or massive hemorrhage occur, Y is recorded as 1, and normal delivery is recorded as 0. The label Y is associated with the feature vector of the full-dimensional spatiotemporal data pool generated by the pregnant woman throughout the perinatal period to form a new training sample pair (X,Y).

[0137] Step S520: Calculate the predicted risk RI using the new training sample pair (X,Y) through the cross-entropy loss function. predict The difference from the actual ending Y;

[0138] Step S530: Using the Adam optimizer, adjust the output layer weights W in the multilayer perceptron in the opposite direction of the gradient. out and bias b out Update;

[0139] Step S540: Based on the collected pregnancy outcome data, the false positive rate and false negative rate under the current threshold {T1,T2} are analyzed by constructing an ROC curve. For T1, if clinical feedback indicates a large number of false negatives, T1 is automatically lowered to expand the monitoring coverage. For T2, the value of T2 is dynamically optimized based on the actual rescue success rate and the use of medical resources to ensure the seriousness of triggering rescue instructions.

[0140] In step S540, the dynamic optimization process for T2 includes:

[0141] Step S541: Obtain operational data within a period of time, and compare the RI index of each case that triggers a level 2 warning with the final pregnancy outcome label Y; for adverse outcome samples, record the distribution curve of the RI index when cases such as eclampsia or fetal distress occur; for normal delivery samples, record the highest peak value of the RI index when normal mothers are subjected to transient fluctuations.

[0142] Step S542: Using an incremental learning algorithm, adjust the model parameters W. out and b out The updated RI output distribution is analyzed, and a new value T2' is sought within the new training samples to minimize the combined loss function of the recall rate and the medical resource waste rate for adverse outcomes. Through localized fitting, the system can incrementally learn based on the physical differences of people in different regions, thereby improving the accuracy of early warning. The recall rate is the proportion of samples with ultimately occurring adverse outcomes that successfully trigger a secondary warning, and the medical resource waste rate is... Rate This refers to the proportion of cases in which a level-two warning is triggered and resource pre-reservation logic (such as locking beds or pre-reserving specific blood types) is executed, and the final pregnancy outcome is normal (label Y=0) or the reserved resources are not actually consumed;

[0143] Step S543: Update the threshold using a smoothing mechanism.

[0144] T2 new =(1-α)·T2 old +αT2 calculated

[0145] Among them, T2 calculated It is the theoretically optimal threshold; α is the learning rate, which ensures that the self-evolution of the evaluation model is robust.

[0146] The comprehensive loss function Loss(T2) in step S542 is expressed as follows:

[0147] Loss(T2)=w1·(1-Recall)+w2·(Wastage Rate )

[0148] Among them, w1 and w2 are weighting factors.

Claims

1. A perinatal whole-cycle intelligent intervention method based on multi-source heterogeneous data fusion, characterized in that, The method comprises the following steps: Step S100, synchronously collecting subjective self-report data of a pregnant woman, real-time physiological data of a wearable device, and clinical examination heterogeneous data, and constructing a full-dimensional space-time data pool; Step S200, establishing a unified time base with gestational weeks and absolute time as double coordinate axes, and extracting feature vectors of data in the full-dimensional space-time data pool according to a fixed time step; Step S300, calculating a real-time global comprehensive risk index RI according to the feature vectors; Step S400, when the RI exceeds a first dynamic threshold T1, triggering a first-level early warning, and when the RI exceeds a second dynamic threshold T2 and a preset critical feature operator is identified, triggering a second-level early warning, generating an admission preparation list, and sending a rescue instruction to a hospital emergency dispatch system.

2. The method of claim 1, wherein, The specific process of step S200 comprises: Step S210, establishing a unified time base with gestational weeks and absolute time as double coordinate axes, and labeling time marks of the collected heterogeneous data on the unified time base; Step S220, downsampling the real-time physiological data of the wearable device, and using forward filling for the subjective self-report data and the hospital interface data; Step S230, at each time step, a feature vector is created containing three types of attributes: V t = [S self (t), W wear (t), C clinical (last_val)].

3. The method of claim 2, wherein, The value process of downsampling the real-time physiological data of the wearable device in step S220 comprises: Step S221, removing all frequency components higher than half of a new sampling rate through a low-pass filter; Step S222, taking values according to a set step M, and retaining the maximum value, the minimum value and the average value in the step.

4. The method of claim 2, wherein, The value process of using forward filling for the subjective self-report data and the hospital interface data in step S220 comprises: Step S223, determining an influence effectiveness window of each data; Step S224, set the step M, in the initial step M start Obtain subjective self-report data and hospital interface data, detect whether the data field is empty in the subsequent time segment {M start+1 ,M start+2 ,...}, if empty and within the validity period, copy the value of the initial time segment M start to the data field of the current time slice; if not empty, set the time segment to the initial time segment T start ; Step S225, calculating a confidence coefficient C(t) of data in each time segment, and if C(t) is less than a threshold, re-measuring the subjective self-report data and the hospital interface data, ， where t is the time difference from the last measurement, Co is the original confidence, T valid is the clinical validity period, λ is the decay factor, and e is the natural logarithm.

5. The method of claim 1, wherein, The specific process of step S300 comprises: Step S310, decoupling the feature vectors into a clinical examination feature subspace, a wearable device physiological feature subspace and a subjective self-report feature subspace; Step S320, construct a cross-attention mechanism network based on clinical guidance, and linearly map the clinical examination features F clinical into a query vector matrix Q c , the physiological features F wear of the wearable device and the subjective self-reported features F self are respectively converted into corresponding key vector matrices K w , K s and value vector matrices V w , V s ; Step S330, respectively calculating scaled dot products of a query vector matrix and two kinds of key vector matrices, and obtaining a cross-attention weight matrix after normalization processing by a Softmax function; Step S340, respectively performing weighted summation on corresponding value vector matrices by using the cross-attention weight matrix, obtaining physiological context features and behavior context features, and performing multi-dimensional feature splicing on the clinical examination features, the physiological context features and the behavior context features to generate a global fusion feature tensor; Step S350, inputting the global fusion feature tensor into a multi-layer perception network to perform nonlinear mapping, and outputting a global comprehensive risk index RI in real time.

6. The method of claim 5, wherein, Q c with K w scaled dot product, and Q c with K s dot product is given by ， ， where d k is the dimension of the bond vector.

7. The method of claim 5, wherein, The method for obtaining the global comprehensive risk index RI by step S350 through the multi-layer perception network comprises ， where F fusion is the global fusion feature tensor, is a nonlinear mapping, W out and b out are the output layer weights and bias learned by the model in the pre-training phase, and σ is an activation function.

8. The method of claim 7, wherein, Output layer weights W out and bias b out The learning process is as follows: Step S351, extracting features in the subjective self-report data, the real-time physiological data and the clinical examination heterogeneous data of an input sample X to form a feature vector, and marking a pregnancy outcome in the input sample as a label Y, and marking an adverse outcome as 1 and otherwise as 0; Step S352, randomly generate output layer weight W out and bias b out ; Step S353, input the feature vector, pass through the attention mechanism and the multi-layer perception calculation, and obtain a predicted risk RI by W out and b out mapping predict ; Step S354, calculating a gap between a predicted risk and a real outcome label Y by using a cross-entropy loss; Step S355, using the chain rule, compute the gradient of the loss function with respect to W out and b out ; Step S356, update W in the opposite direction of the gradient using the Adam optimizer out and b out , ， Loss is a cross-entropy loss function, and η is a learning rate. Step S357, repeat steps S353 to S356 until the loss function no longer decreases.

9. The method of claim 5, wherein, In step S400, the secondary early warning obtains the rescue instruction through the resource prediction algorithm, and the resource prediction algorithm specifically includes: Step S410, receiving a global comprehensive risk index RI, a risk contribution factor identifier F id and the identified criticality feature operators, the risk contribution factor identifier F id is a cross-attention weight matrix, and the criticality feature operators include a fetal distress operator, a severe preeclampsia operator, and a prenatal hemorrhage risk operator. In step S420, a corresponding relationship between risk contribution factor identifiers of different dimensions and medical resource demand types is constructed, wherein the medical resource demand types at least include bed resources B, blood resources L, emergency green channel resources P, and rescue consumable resources E; Step S430, call the preset resource prediction function RP=f(RI, F id , G week ), calculate the demand probability and demand of various medical resources, G week is the gestational age data; In step S440, based on the prediction result generated by the resource prediction algorithm, the pre-occupancy logic marking of the medical resource is executed through the hospital management system (HIS) standard interface; the pre-occupancy logic marking includes: generating a reserved bed identifier in the bed management system and locking the target bed; sending a pre-order instruction of a specific blood type in the blood management system; automatically generating an admission preparation list matched with the risk contribution factor and pushing it to the medical terminal.

Images