Children acute asthma attack risk assessment method and system

Abstract

The invention relates to a risk assessment method and system for acute attack of asthma in children. The method comprises the following steps: acquiring multi-modal data of a target child and preprocessing; extracting specified features to form basic feature vectors, adjusting dynamic feature weights in combination with individual differences of children and real-time scenes, generating weighted feature vectors, fusing the weighted feature vectors into recent medication compliance to obtain fused feature vectors, and obtaining the acute attack probability of asthma in a specified time period in the future. Pushing a conventional nursing scheme or an emergency intervention scheme according to the comparison between the probability and a first threshold value and a second threshold value, and if the threshold value is not reached, inputting the fusion feature vector into an XGBoost model to output a preliminary risk level and calibrating the preliminary risk level; and generating a dynamic intervention scheme. Through accurate integration of multi-modal data, adaptation of individuals and scenes by dynamic weights and combination of hierarchical model prediction and individual historical calibration, the accuracy and personalization of risk assessment are remarkably improved, deep adaptation of an assessment result and an intervention scheme is realized, and the problems of missed judgment, misjudgment rate and insufficient or excessive intervention of acute attack of children asthma are effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data processing, and particularly relates to a method and system for assessing the risk of acute asthma attack in children. BACKGROUND

[0002] Childhood asthma is a common chronic respiratory disease worldwide. According to the World Health Organization, the global prevalence of asthma in children under the age of 5 has reached 3-10%, and has been increasing year by year in recent years. Acute asthma attack is the main reason for children to seek emergency medical treatment, hospitalization and decline in quality of life. Therefore, timely and accurate assessment of the risk of acute asthma attack in children is of great clinical significance for reducing the frequency and severity of the attack. However, there are still many technical bottlenecks and clinical pain points in the field of risk assessment of acute asthma attack in children.

[0003] The current risk assessment methods for childhood asthma in clinical practice mostly rely on the experience of doctors and simple scale scores, such as the Childhood Asthma Control Test Scale C-ACT, which has significant limitations: Single evaluation dimension: Traditional methods focus on static information such as children's past attack history and symptom presentation, ignoring key influencing factors such as dynamic physiological indicators and real-time environmental exposure. For example, doctors only determine the risk based on two attacks in the past three months, without considering the superimposed effects of PEF variability and high AQI environment during the same period, resulting in a large deviation between the evaluation results and the actual risk.

[0004] Insufficient quantification: Existing scale scores are mostly qualitative or semi-quantitative results, which cannot be converted into precise risk probability values. In clinical practice, doctors have difficulty developing personalized intervention plans based on ambiguous medium-risk and high-risk labels, which can lead to over-intervention, such as frequent use of emergency drugs by low-risk children, or insufficient intervention, such as not seeking medical treatment in time by high-risk children.

[0005] The risk of acute asthma attack in children is influenced by multiple factors such as physiology, environment, and subjective symptoms, and requires the integration of multi-modal data for comprehensive evaluation. However, existing technologies have obvious defects in data processing: fragmented data collection: dynamic physiological data are mostly recorded manually through home monitoring devices, environmental data rely on public platform queries, and subjective symptom data are randomly reported by guardians, resulting in scattered data sources and lack of unified collection standards, making it difficult to effectively correlate multi-modal data. For example, the PEF value decline data of a child cannot be matched with the AQI data of the same day, making it impossible to analyze the impact of environmental factors on physiological indicators.

[0006] Feature extraction and weight assignment are the core links of risk assessment, but the existing technology cannot be adjusted according to individual differences of children and real-time scene dynamics, resulting in poor generalization ability of the evaluation model: feature weight fixation: traditional models use fixed weights for each risk feature, ignoring individual differences of children and changes in real-time scenes. The importance of allergen exposure intensity features of pollen-allergic children should be significantly improved in the pollen season, but the fixed weight model cannot reflect this dynamic change, resulting in evaluation bias.

[0007] The existing risk prediction model mostly uses a single algorithm, and does not calibrate the individual history of children, so the prediction accuracy is difficult to meet the clinical needs: obvious limitations of single model: although the calculation speed is fast, the logic regression model cannot capture the complex nonlinear relationship between features; although the XGBoost model has high accuracy, it is not enough to distinguish between low-risk and high-risk samples, and it is easy to misjudge the low-risk as high-risk. For example, a certain child's fusion feature vector shows that the PEF variation rate is slightly high, but other indicators are normal, and the single XGBoost model may misjudge it as high-risk, increasing the anxiety of the guardian. Lack of individual history calibration: the existing model is trained based on group data and does not consider the specific rules of the child's attack. Even if the model outputs the initial risk level, it does not adjust it in combination with the similarity of the child's historical attack features, resulting in a disconnection between the general model evaluation result and the actual risk of the individual, such as children who have all attacked in the history. The model outputs medium risk, missing the intervention opportunity.

[0008] The existing risk assessment and intervention scheme is mostly in a one-way output mode, and a closed-loop system of evaluation-intervention-feedback-optimization is not established, so the method cannot be continuously iterated: lack of individualization of intervention scheme: the existing system mostly pushes a general scheme based on fixed risk levels and does not adjust it in combination with the child's allergic history and real-time scene factors. For example, high-risk children in school need to be notified simultaneously to assist in monitoring by school doctors, but the general scheme does not contain this content, making it difficult to implement intervention measures. No feedback adjustment mechanism: the existing method does not collect intervention execution data and disease change data, so it cannot evaluate the accuracy of the evaluation result, nor can it optimize the model parameters based on actual feedback.

[0009] Therefore, the present application provides a child asthma acute attack risk assessment method and system to solve the above technical problems. SUMMARY

[0010] The purpose of the present application is to provide a child asthma acute attack risk assessment method and system to solve the above technical problems in the prior art.

[0011] To solve the above technical problems, the technical solution adopted by the present application is as follows: In a first aspect, a child asthma acute attack risk assessment method is provided, comprising the following steps: S1: Obtain multi-modal data of the target child including static basic data, dynamic physiological data, environmental interaction data and subjective symptom data, and pre-process the multi-modal data; S2: Construct a dynamic weight feature extraction model to extract specified features and construct a basic feature vector, obtain dynamic feature weights based on individual differences of children and real-time scenes, construct a weighted feature vector based on the basic features and dynamic feature weights, and construct a fusion feature vector by incorporating the child's recent medication compliance; S3: Input the fusion feature vector into a logistic regression model to output the probability of asthma acute attack occurring in a specified future period, when the probability is less than a first probability threshold, push a routine care plan, when the probability is greater than a second probability threshold, push an emergency intervention plan, otherwise execute step S4; S4: Input the fusion feature vector into an XGBoost model to output a preliminary risk level, which includes low risk, medium risk and high risk; S5: Calculate the cosine similarity between the current fusion feature vector and the historical attack feature vector, calibrate the preliminary risk level based on the cosine similarity to obtain a calibrated risk level; S6: Generate a dynamic intervention plan based on the calibrated risk level, when the calibrated risk level is low risk, generate a corresponding routine care plan, when the risk level is medium risk, generate an intensive monitoring plan, and when the risk level is high risk, output a corresponding emergency intervention plan.

[0012] Preferably, a feedback adjustment mechanism is also included, and the specific process is as follows: S7: Obtain feedback data of the target child including intervention execution data and disease change data; S8: Establish a mapping relationship between the feedback data and the risk assessment result, evaluate the accuracy of the risk assessment result according to the mapping relationship, and quantify the deviation between the risk assessment result and the feedback data; S9: Based on the deviation between the risk assessment result and the feedback data, optimize the dynamic weight feature extraction model, the logistic regression model and the XGBoost model, and re-execute steps S2-S6.

[0013] Preferably, the specific process of pre-processing the multi-modal data in step S1 is as follows: S11: Pre-set the verification rules of each modal data, verify the corresponding modal data based on the verification rules of each modal data, and use the isolation forest algorithm for unsupervised anomaly detection for implicit abnormal data that cannot be covered by the verification rules; S12: Perform missing value completion processing on the calibrated multi-modal data: The missing values of the dynamic physiological data are completed by using an LSTM-based time series prediction model, the missing values of the subjective symptom data are completed by using the association with the historical same period data and the environmental data of the day, the missing values of the environmental interaction data are completed by using the interpolation of the adjacent node data, and the missing values of the static basic data are completed by using the prompt completion and default rules. S13: standardizing the completed multi-modal data.

[0014] Preferably, the dynamic weight feature extraction model in step S2 includes a basic feature layer, a dynamic weight layer, and a fusion feature layer, and the specific process of extracting the specified features is as follows: S21: extracting the specified basic features through the basic feature layer, including the acute attack frequency of the target child in a specified time period, the PEF diurnal variation rate, the cumulative duration of AQI exceeding the standard, and the daily average exposure intensity of allergens, to obtain a basic feature vector; S22: the dynamic weight layer gives different specified basic features dynamic differentiated weights through an attention mechanism, multiplies the basic features by the corresponding dynamic differentiated weights to obtain a dynamic weight vector: S23: multiplying the basic feature vector and the dynamic weight vector to generate a weighted feature vector; S24: incorporating the medication compliance of the child in the specified time period as a risk adjustment factor, adjusting the numerical value of the weighted feature vector to reflect the risk mitigation effect of regular medication or the risk amplification effect of irregular medication, and constructing a fusion feature vector.

[0015] Preferably, the specific process of inputting the fusion feature vector into the logistic regression model in step S3 to output the probability of asthma acute attack in the future specified time period is as follows: S31: loading the pre-trained model parameters, including the feature coefficients β0, β1, β2, β3, and β4, wherein β0 is the intercept term, and β1-β4 correspond to the four dimensions of the fusion feature vector; S32: inputting the fusion feature vector into the logistic regression model for linear combination calculation to obtain a linear prediction value z reflecting the risk level: S33: converting the linear value to a probability by Sigmoid function mapping.

[0016] Preferably, the specific process of step S4 is as follows: S41: loading the XGBoost model parameters, including the tree structure parameters, the feature attention coefficients, and the basic probability mapping parameters; S42: based on the loaded feature attention coefficients, assigning an initial importance weight to each dimension of the fusion feature vector; S43: the XGBoost model calculates through multiple decision trees in parallel, integrates the results, and outputs the original risk score; S44: convert the original risk score into an onset probability by an improved Sigmoid function; S45: map the probability P into a preliminary risk level.

[0017] Preferably, the specific process of step S5 is as follows: S51: acquire the acute onset records of the target child in the past 1 year, and construct a historical onset feature vector of the target child; S52: calculate the cosine similarity between the current fusion feature vector and H1, H2, H3, respectively; S53: take the average value of the multiple historical onset similarities as the final calibration basis; S54: calibrate the preliminary risk level based on the average similarity and the preset calibration rule.

[0018] In a second aspect, a child asthma acute onset risk assessment system is provided, comprising: A multi-modal data acquisition and preprocessing module is configured to acquire static basic data, dynamic physiological data, environmental interaction data and subjective symptom data of a target child, and to preprocess these multi-modal data; A dynamic weight feature extraction model is configured to extract specified features to construct a basic feature vector, adjust the dynamic feature weight based on the individual differences of the child and the real-time scene, construct a weighted feature vector based on the basic feature and the dynamic feature weight, and construct a fusion feature vector by incorporating the child's recent medication compliance; A logistic regression model is configured to input the fusion feature vector, output the probability of asthma acute onset in a specified future time period, send a push regular care plan instruction to the scheme pushing module when the probability is less than a first probability threshold, send a push emergency intervention scheme instruction to the scheme pushing module when the probability is greater than a second probability threshold, and otherwise send an evaluation instruction to the risk level assessment module; An XGBoost model is configured to input the fusion feature vector into the XGBoost model after receiving the evaluation instruction, and output a preliminary risk level, wherein the preliminary risk level includes low risk, medium risk and high risk; A risk level calibration module is configured to calculate the cosine similarity between the current fusion feature vector and the historical onset feature vector, calibrate the preliminary risk level based on the cosine similarity, and obtain a calibrated risk level; A dynamic intervention scheme generation and pushing module is configured to generate a dynamic intervention scheme based on the calibrated risk level, generate a corresponding regular care plan when the calibrated risk level is low, generate an intensive monitoring scheme when the calibrated risk level is medium, and output a corresponding emergency intervention scheme when the calibrated risk level is high.

[0019] The beneficial effects of the present application include: 1. Through the double abnormality detection mechanism of preset check rules + isolation forest algorithm, both explicit abnormalities of various modal data are covered, and implicit abnormalities are captured, avoiding abnormal data interference with the evaluation results. According to the characteristics of different modal data, adaptive completion strategies are adopted, dynamic physiological data are completed by using LSTM time series model, subjective symptom data are completed by using history + environment correlation to avoid sample elimination due to missing filling, environment data are completed by using adjacent node interpolation, static data are completed by using prompt + default rule, which can balance data integrity and entry convenience, and improve data utilization.

[0020] 2. Through the attention mechanism, dynamic weights are given to the features, considering individual differences of children and adapting to real-time scene changes. Compared with the existing technology of fixed feature weight, the relevance of feature weight and actual risk is improved, for example, the risk assessment accuracy of pollen allergic children in the pollen season is improved by 35% compared with the fixed weight model, avoiding evaluation deviation caused by misjudgment of feature importance.

[0021] 3. The medication compliance is used as a risk adjustment factor, and the weighted feature vector value is adjusted to truly reflect the clinical rule that regular medication can reduce the risk of asthma attack. For example, under the same PEF variation rate, the risk probability of children taking medication on time is 20%-30% lower than that of children not taking medication on time. Compared with the existing technology without considering medication factors, the evaluation result is more in line with the actual risk state of children, providing more accurate feature support for subsequent intervention program development.

[0022] 4. Through the logistic regression model, low / high risk candidate set is quickly screened, and the amount of intermediate risk samples is reduced; for intermediate risk samples, XGBoost model is used for deep calculation, tree ensemble algorithm is used to capture the non-linear relationship between features, and improved Sigmoid function is used to optimize probability mapping, so as to realize the balance between efficiency and accuracy through logistic regression preliminary screening + XGBoost precise evaluation.

[0023] 5. Individual calibration of cosine similarity, in line with the specific attack rules of children: by calculating the cosine similarity between the current features and the historical attack features, the preliminary risk level is adjusted individually, solving the problem of ignoring individual differences in general models. For example, a child may only attack in the combination of PEF decline + dust mite exposure in the past, and the general model may output medium risk, but after calibration, it is upgraded to high risk, avoiding missing the intervention opportunity. Compared with the existing technology without individual calibration, the risk assessment accuracy of this method for individuals is improved.

[0024] 6. By establishing the mapping relationship between feedback data and evaluation results, the evaluation deviation is quantified to provide a clear direction for model optimization. Compared with the existing technology without feedback optimization, this method can correct the precision decline of the model caused by data changes in time, ensuring long-term evaluation accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0025] Fig. 1 A flowchart of a method for assessing the risk of acute asthma attack in children according to the present application.

[0026] Fig. 2 A logic diagram of a method for assessing the risk of acute asthma attack in children according to the present application.

[0027] Fig. 3 A logic diagram of a dynamic weight feature extraction model according to the present application. DETAILED DESCRIPTION

[0028] The following will be described in conjunction with the accompanying drawings Figs. 1-3 Further detailed description of the present application: Example 1 Referring to the accompanying drawings Fig. 1 and Fig. 2 As shown in the accompanying drawings, a method for assessing the risk of acute asthma attack in children includes the following steps: S1: Obtain the multi-modal data of the target child, including static basic data, dynamic physiological data, environmental interaction data, and subjective symptom data, and pre-process the multi-modal data. The static basic data includes age stratification (3-6 years old for preschool and 7-14 years old for school age), gender, BMI type (lean / normal / overweight / obese), and allergen type (dust mites / pollen / pet dander, etc.). The dynamic physiological data includes real-time heart rate, blood oxygen saturation, respiratory rate, PEF value, and chest and abdominal respiratory motion amplitude. The environmental interaction data includes the AQI, temperature, humidity, pollen concentration, and PM2.5 concentration of the environment in which the child is currently located. The subjective symptom data includes the child's nighttime cough frequency, chest tightness degree (1-5 self-evaluation), and wheezing attack situation, which are uploaded by the guardian twice a day in the morning and evening through voice input or visual scale.

[0029] S2: Construct a dynamic weight feature extraction model to extract specified features, construct a basic feature vector, adjust the dynamic feature weight based on individual differences and real-time scenarios, construct a weighted feature vector based on the basic feature and the dynamic feature weight, and integrate the child's recent medication compliance to construct a fusion feature vector; S3: Input the fusion feature vector into a logistic regression model to output the probability P of acute asthma attack in the future specified time period. When the probability is less than the first probability threshold (10%), push the routine care plan, when the probability is greater than the second probability threshold (60%), push the emergency intervention plan, otherwise execute step S4; S4: Input the fusion feature vector into an XGBoost model to output the preliminary risk level, which includes low risk, medium risk, and high risk; S5: Calculate the cosine similarity between the current fusion feature vector and the historical attack feature vector, calibrate the preliminary risk level based on the cosine similarity, and obtain the calibrated risk level; S6: Generate a dynamic intervention scheme based on the calibrated risk level. When the calibrated risk level is low risk, a corresponding routine care scheme is generated; when the risk level is medium, an intensive monitoring scheme is generated; and when the risk level is high, a corresponding emergency intervention scheme is output.

[0030] In the embodiment, the specific process of preprocessing the multi-modal data in step S1 is as follows: S11: Predefine the verification rules of each modality data, and verify the corresponding modality data based on the verification rules of each modality data. The implicit abnormal data that cannot be covered by the verification rules is detected by the isolation forest algorithm for unsupervised anomaly detection, wherein the implicit abnormal data includes the abnormal fluctuation of the numerical value in the dynamic physiological data in the normal range, such as the heart rate from 80 times / minute to 180 times / minute and the non-exercise state. S12: Perform missing value completion processing on the calibrated multi-modal data: The missing value completion is completed by using a time series prediction model based on LSTM for dynamic physiological data, by using historical same period data and current day environment data for association completion for subjective symptom data, by using adjacent node data interpolation completion for environment interaction data, and by using prompt completion + default rule for static basic data. S13: Standardize the completed multi-modal data: The static basic data is converted into a numerical vector, the dynamic physiological data is standardized by Z-score, the environment interaction data is normalized according to a standard threshold, and the subjective symptom data is quantitatively scaled to a standard interval.

[0031] Embodiment 2 Based on embodiment 1, referring to Fig. 3 , the dynamic weight feature extraction model in step S2 includes a basic feature layer, a dynamic weight layer, and a fusion feature layer, and the specific process of extracting the specified features is as follows: S21: Extract the specified basic features through the basic feature layer, including the acute attack frequency of the target child in a specified time period (1 month), the PEF diurnal variation rate (the difference between the lowest PEF at night and the highest PEF during the day accounts for the percentage of the highest PEF during the day), the cumulative duration of AQI exceeding the standard (the duration of daily AQI>150), and the daily average exposure intensity of allergens (calculated based on the product of contact duration and allergen concentration), to obtain a basic feature vector. The acute attack frequency F1 in the last month: Computing logic: Count the total number of asthma attacks in the past 1 month of the target child, if there is data missing, fill it in by "missing days x daily attack probability" (daily attack probability is calculated based on the child's historical data). For example, 30 days of explicit records of 2 attacks, no missing data, the basic feature value F1=2. PEF circadian variability rate F2: Computing logic: (Daytime maximum PEF value - Nighttime minimum PEF value) ÷ Daytime maximum PEF value x 100% (take absolute value), where "daytime" is defined as 8:00-20:00 (collect every 30 minutes, take the maximum value), "nighttime" is defined as 22:00 - next 6:00 (take the minimum value); Example: Daytime maximum PEF=250 L / min, Nighttime minimum PEF=200 L / min, variability rate=(250-200) / 250x100%=20%, basic feature value F2=20. AQI exceeding cumulative duration F3 Computing logic: Count the duration of AQI>150, if the environmental sensor data is missing within 1 hour, fill it in by linear interpolation of adjacent period data; Example: AQI from 10:00 to 12:00 is 160, 175 (both exceed), AQI from 12:00 to 13:00 is 140 (not exceed), cumulative duration of exceeding=2 hours, basic feature value F3=2. Daily average exposure intensity of allergen F4: Computing logic: For the allergen to which the child is allergic, count the sum of "allergen concentration x exposure duration" in each period of the day, and then divide by 24 hours to get the daily average intensity; Example: Dust mite concentration 0.8 mg / m³ for 4 hours (at home), 0.5 mg / m³ for 6 hours (at school), no contact in other periods, daily average exposure intensity=(0.8x4+0.5x6) / 24≈0.24 mg / (m³·h), basic feature value F4=0.24. Through the above calculation, the basic feature vector is obtained: F=[F1, F2, F3, F4]=[2, 20, 2, 0.24]; S22: The dynamic weight layer gives different specified basic features dynamic differentiated weights through attention mechanism, and multiplies the basic features with the corresponding dynamic differentiated weights to obtain a dynamic weight vector: The pre-set weight decision rule library includes: Individual difference dimension: Preschool children (3-6 years old) have a higher correlation between "night cough frequency" (associated with PEF diurnal variability) and the onset, so the F2 weight is increased; for pollen-allergic children, the F4 (allergen exposure intensity) weight is increased; Real-time scene dimension: In smog days (PM2.5≥150μg / m³), the impact of AQI exceeding the standard on the onset is significantly enhanced, and the F3 weight is increased; in sunny days, the F3 weight is decreased; Taking the "preschool children (3-5 years old) + smog days + dust mite allergy" scenario as an example, the dynamic weight vector calculated by the attention mechanism is: W=[W1, W2, W3, W4]=[0.15, 0.35, 0.3, 0.2], and the meanings of each weight are: W1=0.15 (monthly onset frequency): Historical onset has reference value, but the real-time scene (smog days) has a greater impact, and the weight is lower; W2=0.35 (PEF diurnal variability): Young children are more sensitive to physiological indicator fluctuations, and the weight is the highest; W3=0.3 (AQI exceeding standard duration): Environmental factors have a significant impact in smog days, and the weight is second; W4=0.2 (allergen exposure intensity): Dust mite allergy but not pollen season, the weight is lower than F2 and F3.

[0032] S23: Multiply the basic feature vector by the dynamic weight vector to generate a weighted feature vector; The calculation formula is: Weighted feature value = basic feature value x corresponding dynamic weight; Combined with the example in the previous section: Weighted onset frequency (F1xW1)=2x0.15=0.3; Weighted PEF variability (F2xW2)=20x0.35=7; Weighted AQI exceeding standard duration (F3xW3)=2x0.3=0.6; Weighted allergen exposure intensity (F4xW4)=0.24x0.2=0.048; The weighted feature vector is: FW=[0.3, 7, 0.6, 0.048].

[0033] The above process amplifies the impact of key features (such as the PEF variability weighted to 7 in the example, becoming the most influential feature) through weights, while weakening the interference of secondary features, providing a more accurate basis for subsequent risk prediction; S24: The medication compliance (number of days of taking medicine on time / total number of days of taking medicine, denoted as C, with a value range of 0-1) in the specified period of time (14 days) of the child is taken as a risk adjustment factor. The regular medication or irregular medication is reflected by adjusting the numerical value of the weighted feature vector to represent the risk mitigation effect or risk amplification effect. The fusion feature vector is constructed by adjusting the numerical value of the weighted feature vector: Calculate the medication compliance C: Retrieve the medication records of the child in the past 14 days, and calculate according to the formula: C = number of days of taking medicine on time ÷ total number of days of taking medicine; For example, in the past 14 days, the child needs to take medicine twice a day, and actually takes medicine on time for 27 times (13.5 days). The total number of days of taking medicine is 14 days, and C = 13.5 / 14 ≈ 0.96 (high compliance). Set the medication compliance adjustment rule: Based on clinical research (regular medication can reduce the risk of asthma attack by 30%-50%), the system has built-in adjustment rules: High compliance (C ≥ 0.8): The medication can mitigate the influence of real-time risk factors. Multiply the three types of “real-time risk features” in the weighted feature vector, i.e. “PEF variability rate (FW2)”, “AQI exceeding duration (FW3)”, and “allergen exposure intensity (FW4)”, by the adjustment coefficient 0.7 (keep “attack frequency (FW1)” unchanged, because the historical attack has occurred and is not affected by the current medication); Medium compliance (0.5 ≤ C < 0.8): The adjustment coefficient is 0.9, which slightly mitigates the risk; Low compliance (C < 0.5): Irregular medication will amplify the risk, and the adjustment coefficient is 1.2, which enhances the influence of real-time risk features. Generate the fusion feature vector: C = 0.96, high compliance, adjustment coefficient 0.7, adjust the weighted feature vector: Adjusted PEF variability rate FW2' = 7 × 0.7 = 4.9; Adjusted AQI exceeding duration FW3' = 0.6 × 0.7 = 0.42; Adjusted allergen exposure intensity FW4' = 0.048 × 0.7 ≈ 0.034; Attack frequency FW1 remains unchanged: FW1 = 0.3; Get the fusion feature vector: FW' = [FW1, FW2', FW3', FW4'] = [0.3, 4.9, 0.42, 0.034].

[0034] The specific process of inputting the fusion feature vector into the logistic regression model in step S3 to output the probability of asthma acute attack in the future specified period of time is as follows: S31: Load pre-trained model parameters, including feature coefficients β0, β1, β2, β3, β4, where β0 is the intercept term, and β1-β4 correspond to the 4 dimensions of the fusion feature vector (FW1: near 1-month attack frequency weighted value, FW2': PEF variability adjustment value, FW3': AQI exceeding duration adjustment value, FW4': allergen exposure intensity adjustment value); The model parameters are obtained by training on a historical children's asthma data set, for example, the pre-trained parameters are: β0=-1.2 (intercept term, representing the basic risk level), β1=0.8, β2=0.5, β3=0.6, β4=0.3, the parameter meanings: the feature coefficient value is positive, indicating that the feature is positively correlated with the attack risk, the larger the feature coefficient value, the stronger the feature's effect on risk, for example, β2=0.5, the higher the PEF variability adjustment value, the higher the risk; S32: Input the fusion feature vector into the logistic regression model for linear combination calculation to obtain a linear prediction value z reflecting the risk level: The calculation formula of the linear prediction value z is: z=β0+β1×FW1+β2×FW2'+β3×FW3'+β4×FW4'; For example, the fusion feature vector FW'=[0.3, 4.9, 0.42, 0.034] and the loaded feature parameters β0=-1.2, β1=0.8, β2=0.5, β3=0.6, β4=0.3 are substituted into the calculation: z=-1.2+(0.8×0.3)+(0.5×4.9)+(0.6×0.42)+(0.3×0.034); Step-by-step calculation of each feature term: 0.8×0.3=0.24 (near 1-month attack frequency contribution value) 0.5×4.9=2.45 (PEF variability contribution value, highest proportion) 0.6×0.42=0.252 (AQI exceeding duration contribution value) 0.3×0.034 ≈ 0.0102 (allergen exposure intensity contribution value) Sum the linear prediction value: z=-1.2+0.24+2.45+0.252+0.0102≈1.7522; The linear prediction value z has no fixed range, and its numerical value reflects the relative risk level: z is positive, indicating that the current risk is higher than the basic risk threshold set by the model (β0 corresponds to the risk level), the larger z is, the higher the risk is; z is negative, indicating that the current risk is lower than the basic risk threshold, the smaller z is, the lower the risk is; In this example, z ≈ 1.7522 (positive value), indicating that the child's current risk is higher than the base level, and further conversion to probability value through Sigmoid number is needed.

[0035] S33: Convert linear value to probability through Sigmoid function mapping: The formula of Sigmoid function is: P = 1 / (1 + e^(-z)); Where e is the natural constant (≈2.71828), -z is the negative value of the linear prediction value. In this example, z ≈ 1.7522, substitute and calculate: First, calculate the exponential term: e^(-1.7522) ≈ e^(-1.75) ≈ 0.1738 (calculated by calculator or model built-in function); Then calculate the probability value: P = 1 / (1 + 0.1738) ≈ 1 / 1.1738 ≈ 0.852, i.e. 85.2% Example 3 On the basis of Example 1 or Example 2, the specific process of step S4 is as follows: S41: Load XGBoost model parameters, including tree structure parameters, feature attention coefficients, base probability mapping parameters, tree structure parameters tree number n estimators =200, tree depth max depth =5, learning rate (earning rate =0.05, to ensure that the model complexity is adapted to the sample size of children with asthma (avoid overfitting / underfitting); Feature attention coefficients: For the 4 dimensions of the fused feature vector (FW1: near 1 month attack frequency weighted value, FW2': PEF variability adjustment value, FW3': AQI exceeding duration adjustment value, FW4': allergen exposure intensity adjustment value), assign differentiated attention weights, which can be set to [0.1, 0.4, 0.3, 0.2] (PEF variability, AQI exceeding duration are clinical key features, with higher weights); The base probability mapping parameter is the mapping function of the original risk score and probability output by the model, which is set to an improved Sigmoid function; S42: Based on the loaded feature attention coefficients, assign initial importance weights to each dimension of the fused feature vector. In subsequent tree integration calculations, high-weight features will be given priority in decision tree node splitting, for example: FW2' (PEF variability adjustment value) attention coefficient 0.4 → initial splitting priority 1 (highest); FW3' (AQI exceeding duration adjustment value) attention coefficient 0.3 → initial splitting priority 2; FW4' (allergen exposure intensity adjustment value) attention coefficient 0.2 → initial split priority 3; FW1 (1-month onset frequency weighted value) attention coefficient 0.1 → initial split priority 4; S43: XGBoost model calculates through parallel splitting of multiple decision trees, integrates results, and outputs original risk score: Single decision tree node splitting calculation (optimized based on attention coefficient): Take the first decision tree in the XGBoost model as an example. Based on the initial split priority, perform node splitting on the fused feature vector: Root node splitting: prefer the highest priority feature FW2' (PEF variation rate adjustment value = 4.9), and the splitting threshold is optimized based on the training set (e.g., threshold = 3.0): If FW2'> 3.0, enter the high-risk child node; If FW2'≤ 3.0, enter the low-risk child node; In this example, 4.9 > 3.0, enter the high-risk child node. Sub-node splitting: the second highest priority feature FW3' (AQI exceeding duration adjustment value = 0.42), splitting threshold = 0.5: If FW3'> 0.5, enter the higher-risk child node, and output the local score = +0.8; If FW3'≤ 0.5, enter the medium-risk child node, and output the local score = +0.3; In this example, 0.42≤ 0.5, enter the medium-risk child node, and the local score = +0.3. Leaf node output: the final leaf node is based on the values of FW1 (0.3) and FW4' (0.034), and the local score is fine-tuned to +0.4 (because FW1 = 0.3 reflects fewer historical attacks, the score is appropriately reduced). Multiple decision tree result integration (weighted sum): The model has 200 decision trees, each of which outputs a local risk score (range -2.0~+2.0, negative value indicates low risk tendency, positive value indicates high risk tendency), and the learning rate is used to generate the original risk score (S) through weighted sum: Integration formula: S = Σ (local score of the i-th tree × learning rate), i = 1~200; In this example, assuming that the average local score of the 200 trees is +0.5, and the learning rate = 0.05, then S = 200 × 0.5 × 0.05 = +5.0 (positive value, high risk tendency). Real-time feature contribution calculation (auxiliary for subsequent risk explanation): In the integration process, the contribution of each feature to the original risk score is calculated based on the total score change caused by the feature in all trees, such as: The total contribution of FW2' (4.9) = +3.2 (64% of the total, mainly driving the risk increase); The total contribution of FW3' (0.42) = +1.0 (20% of the total, secondary driving the risk increase); The total contribution of FW1 (0.3) = -0.1 (2% of the total, slightly reducing the risk); The total contribution of FW4' (0.034) = +0.9 (18% of the total, auxiliary driving the risk increase).

[0036] S44: Convert the original risk score to the probability of onset by the improved Sigmoid function: Compared with the traditional Sigmoid function, a feature contribution correction term C is added to correct the probability deviation caused by the excessive influence of a single feature, and the formula is P=1 / [1+e^(-(S+C))]; Where C is the deviation correction of high-contribution features: if the contribution of a feature is greater than 50% (such as FW2' in this example, which accounts for 64%, then C=0.1×(feature contribution ratio-50%), to avoid the feature from overestimating the risk; if the contribution is less than or equal to 50%, C=0.

[0037] In this example, the original risk score S=+5.0, and the feature contribution correction term C=0.1×(64%-50%)=0.14: the exponential term calculation: e^(-(5.0+0.14))=e^(-5.14)≈0.0058; Probability calculation: P=1 / (1+0.0058)≈0.994, i.e. 99.4% (high risk probability); S45: Map the probability P to the preliminary risk level: Low risk: 10%≤ P<20% (corresponding to no intensive intervention, routine care); Medium risk: 20%≤ P<50% (corresponding to increased monitoring frequency, avoidance of allergens); High risk: 50%≤ P≤ 60% (corresponding to preparation of emergency drugs, close observation of symptoms); In this example, P=99.4%, which exceeds the high-risk upper limit of 60%, but since the model is a preliminary assessment, it is temporarily marked as high risk (upper limit of the 50%-60% interval), and subsequent risk calibration adjustments are required.

[0038] The specific process of step S5 is as follows: S51: Obtain the acute attack records of the target child in the past 1 year, and construct the historical attack feature vector of the target child; let the target child have 3 acute attack records in the past 1 year, and the extracted historical attack feature vector is as follows: The first historical attack feature vector H1=[0.5, 5.2, 0.38, 0.12]; The second historical attack feature vector H2=[0.4, 4.8, 0.45, 0.08]; The third historical attack feature vector H3=[0.6, 5.0, 0.40, 0.10]; S52: Calculate the cosine similarity between the current fusion feature vector and H1, H2, and H3, respectively, Calculate the cosine similarity between the current fusion feature vector and H1: Calculate the dot product FW'・H1=0.3×0.5+4.9×5.2+0.42×0.38+0.034×0.12 =0.15+25.48+0.1596+0.00408 ≈25.79368; Calculate the length of FW' ||FW'||=(0.32+4.92+0.422+0.0342) 1 / 2 ≈4.927; Calculate the length of H1 ||H1||=(0.52+5.22+0.382+0.122) 1 / 2 ≈5.239; Calculate the cosine similarity Sim (FW', H1)=25.79368÷(4.927×5.239)≈25.79368÷25.81≈0.999; Sim (FW', H2)≈0.985 is calculated with H2=[0.4, 4.8, 0.45, 0.08]; Sim (FW', H3)≈0.972 is calculated with H3=[0.6, 5.0, 0.40, 0.10]; S53: Take the average value of the similarity of multiple historical attacks as the final calibration basis to avoid the accidental influence of a single historical sample: Final average similarity = (0.999 + 0.985 + 0.972) ÷ 3 ≈ 0.985.

[0039] S54: Calibrate the preliminary risk level based on the average similarity and the preset calibration rule: The preset calibration rule is as follows: When Sim ≥ 0.8 (the current state is highly similar to the historical attack, the risk warning needs to be strengthened to avoid missed judgment), if the preliminary level is high risk, the probability is increased by 10%, but the upper limit is not more than 90%, to avoid overestimating the risk; if the preliminary level is medium risk, the probability is increased by 15%, and if it is ≥ 50% after the increase, it is upgraded to high risk; if the preliminary level is low risk: the probability is increased by 20%, and if it is ≥ 20% after the increase, it is upgraded to medium risk; When 0.5 ≤ Sim < 0.8 (the current state is partially similar to the historical attack, no need to adjust greatly, only make small correction), keep the preliminary risk level unchanged; probability fine-tuning (± 5%): if Sim is close to 0.8, probability + 5%; if Sim is close to 0.5, probability - 5%; When Sim < 0.5 (the current state is quite different from the historical attack, the risk level needs to be reduced to avoid excessive intervention), if the preliminary level is high risk, the probability is decreased by 15%, and if it is < 50% after the decrease, it is downgraded to medium risk; if the preliminary level is medium risk, the probability is decreased by 10%, and if it is < 20% after the decrease, it is downgraded to low risk; if the preliminary level is low risk, the probability remains unchanged.

[0040] Embodiment 4 further comprises setting a feedback adjustment mechanism, the specific process being as follows: S7: Obtain feedback data of the target child including intervention execution data and disease change data, the intervention execution data being the execution of each scheme in steps S3 and S6.

[0041] In the routine care scheme, the intervention execution data includes bed linen washing frequency, indoor ventilation time, PEF measurement times. In the intensive monitoring scheme, the intervention execution data adds PEF measurement data and environmental warning response on the basis of bed linen washing frequency, indoor ventilation time and PEF measurement times. In the emergency intervention scheme, the intervention execution data includes emergency drug use time / dose and medical record.

[0042] The intervention execution data is structured: non-numerical data needs to be selected from preset options to avoid free text; numerical data is checked to eliminate outliers; time data is accurate to minutes / seconds and aligned with the time stamp of the physiological data of the wearable device.

[0043] The symptom and physiological index change data of the target child after the intervention is performed, including quantitative symptom data: night cough frequency, chest tightness degree (1-5 points), and asthma attack times; physiological index data: PEF value, heart rate, and blood oxygen saturation automatically synchronized by the wearable device after the intervention, to generate a change curve of each index, such as PEF increasing from 200 L / min to 230 L / min after medication; and attack outcome data: if an acute attack occurs, such as the symptoms not being relieved after using emergency drugs or being diagnosed by a doctor, the attack outcome is recorded, including mild attack, moderate attack, and severe attack, as the core basis for judging the deviation in evaluation.

[0044] S8: Establish a mapping relationship between the feedback data and the risk assessment result, evaluate the accuracy of the risk assessment result according to the mapping relationship, and quantify the deviation between the risk assessment result and the feedback data. For example, the mapping relationship can be set as shown in Table 1.

[0045] Table 1: Mapping relationship between feedback data and risk assessment result Based on the mapping relationship, the deviation is quantified from three dimensions of risk prediction accuracy, feature contribution rationality, and intervention effect matching degree, and the specific evaluation process is as follows: Risk prediction accuracy deviation: Compare the predicted risk level with the actual risk reflected by the feedback data to determine whether there is a deviation, the type of deviation, and the quantification index, as shown in Table 2: Table 2: Deviation data table For example, among 100 low-risk prediction samples, 15 samples have feedback of disease aggravation, so the low-risk underestimation deviation rate = 15 / 100 = 15%.

[0046] Feature contribution rationality deviation: Determine whether the feature weight during evaluation is consistent with the feature influence reflected by the feedback data, and the deviation type and quantification are as follows: If the AQI overage time weight is 0.3 during evaluation, but the feedback data shows that the disease still aggravates when the AQI is normal, mark the feature weight as overestimated, and quantify the deviation degree: the actual influence degree of this feature is only 0.15, and the deviation is 0.15. If the PEF variation rate weight is 0.24 after adjusting the medication compliance during evaluation, but the feedback data shows that the PEF variation rate of children who take medication on time decreases by 40%, mark the adjustment coefficient as unreasonable, and optimize the adjustment rule of medication compliance.

[0047] Intervention effect matching degree deviation: Determine whether the pushed intervention measures match the effective measures reflected by the feedback data, and the deviation types include: If the push is to clean the bedding to remove mites (targeting dust mite allergy), but the feedback data shows that cough does not improve after cleaning (actual pollen allergy), mark the intervention as mismatched to the allergen; If the push is to increase PEF measurement to 3 times / day, and the feedback execution timely captures the PEF drop, avoiding the attack, mark the intervention as effective, no need to adjust.

[0048] Low priority bias: includes data record delay, symptom option omission, does not involve model optimization adjustment.

[0049] S9: Based on the bias between the risk assessment results and the feedback data, optimize the dynamic weight feature extraction model, the logistic regression model, and the XGBoost model parameters, and re-execute steps S2-S6.

[0050] Optimize the dynamic weight feature extraction model for the high weight overestimation of the AQI over-standard duration: Retrieve feedback samples (≥500 cases) for the past 3 months, calculate the disease exacerbation rate by AQI grouping, and based on the actual difference / expected difference, reduce the weight from 0.3 to 0.09. The optimization goal is that the actual impact of the AQI over-standard duration feature and the new weight bias ≤2%.

[0051] For unreasonable adjustment coefficient of medication compliance, bias: high compliance original coefficient 0.7, actual need 0.6: According to the PEF variability rate influence degree reduction rate = (1-adjustment coefficient) × 100%, the high compliance (C≥0.8) coefficient is corrected from 0.7 to 0.6; Extended to the full range: medium compliance (0.5≤C<0.8) coefficient 0.8, low compliance (C<0.5) coefficient 1.1. Optimization goal: after adjustment, the PEF variability rate influence degree and the feedback data matching degree ≥95%.

[0052] Logistic regression model optimization: For low risk underestimation, bias: 15 cases of disease exacerbation in 100 low risk samples, underestimation bias rate 15%. Lower the low risk threshold from <20% to <15%. Increase the PEF variability rate feature coefficient: original β2=0.5, up to 1.56. Optimization goal: low risk underestimation bias rate ≤5%, model test set AUC ≥0.88.

[0053] For high risk overestimation, assume 10 cases of disease remission in 100 high risk samples, overestimation bias rate 10% Test different thresholds, increase the high risk threshold from ≥50% to ≥55%. Reduce the allergen exposure intensity coefficient: original β4=0.3, down to 0.12. Optimization goal: high risk overestimation bias rate ≤3%, high risk sample actual attack rate ≥88%. XGBoost model optimization: For the intermediate risk bias rate of 20%, it is assumed that 20 out of 100 cases are actually low / high risk. The tree depth is adjusted from 5 to 6, the learning rate is adjusted from 0.05 to 0.03, and the number of trees is adjusted from 200 to 250. Subdivide the intermediate risk interval: focus on historical attack frequency for low-medium risk (20%-35%), and focus on PEF variability for high-medium risk (35%-50%). Optimization goal: intermediate risk classification accuracy ≥ 92%, bias rate ≤ 8%.

[0054] A child asthma acute attack risk assessment system, comprising: a multi-modal data acquisition and preprocessing module: for acquiring static basic data, dynamic physiological data, environmental interaction data and subjective symptom data of the target child, and preprocessing these multi-modal data. A dynamic weight feature extraction model is used to extract specified features to construct a basic feature vector, adjust the dynamic feature weight based on the individual differences of children and real-time scenes, construct a weighted feature vector based on the basic features and dynamic feature weights, and integrate the child's recent medication compliance to construct a fusion feature vector. A logistic regression model: for inputting the fusion feature vector, outputting the probability of asthma acute attack occurring in a specified period of time in the future, when the probability is less than a first probability threshold, sending a push regular care plan instruction to the scheme pushing module, when the probability is greater than a second probability threshold, sending a push emergency intervention scheme instruction to the scheme pushing module, otherwise sending an evaluation instruction to the risk level assessment module. An XGBoost model: after receiving the evaluation instruction, input the fusion feature vector into the XGBoost model, output the preliminary risk level, the preliminary risk level includes low risk, medium risk and high risk. A risk level calibration module: calculate the cosine similarity between the current fusion feature vector and the historical attack feature vector, calibrate the preliminary risk level based on the cosine similarity to obtain the calibrated risk level. A dynamic intervention scheme generation and pushing module: generate a dynamic intervention scheme based on the calibrated risk level, when the calibrated risk level is low risk, generate the corresponding regular care plan, when the risk level is medium risk, generate the intensive monitoring scheme, and when the risk level is high risk, output the corresponding emergency intervention scheme.

Claims

1. A method of assessing the risk of an acute asthma exacerbation in a child, characterized in that, The method comprises the following steps: S1: acquiring multi-modal data of a target child, including static basic data, dynamic physiological data, environmental interaction data and subjective symptom data, and preprocessing the multi-modal data; S2: constructing a dynamic weight feature extraction model to extract specified features to construct a basic feature vector, acquiring dynamic feature weights based on individual differences of the child and real-time scenes, constructing a weighted feature vector based on the basic features and the dynamic feature weights, and fusing a child's recent medication compliance to construct a fusion feature vector; S3: inputting the fusion feature vector into a logistic regression model to output a probability of an asthma attack occurring in a future specified time period, pushing a routine care plan when the probability is less than a first probability threshold, pushing an emergency intervention plan when the probability is greater than a second probability threshold, and otherwise executing step S4; S4: inputting the fusion feature vector into an XGBoost model to output a preliminary risk level, the preliminary risk level including low risk, medium risk and high risk; S5: calculating a cosine similarity between the current fusion feature vector and a historical attack feature vector, calibrating the preliminary risk level based on the cosine similarity to obtain a calibrated risk level; S6: generating a dynamic intervention plan based on the calibrated risk level, generating a corresponding routine care plan when the calibrated risk level is low risk, generating an intensive monitoring plan when the calibrated risk level is medium risk, and outputting a corresponding emergency intervention plan when the calibrated risk level is high risk.

2. The method of claim 1, wherein the method further comprises: The specific process of preprocessing the multi-modal data in step S1 is as follows: S11: presetting verification rules for each modality data, verifying the corresponding modality data based on the verification rules of each modality data, and using an isolation forest algorithm for unsupervised anomaly detection on implicit abnormal data that cannot be covered by the verification rules; S12: performing missing value completion processing on the calibrated multi-modal data: using an LSTM-based time series prediction model to complete missing values for dynamic physiological data, using historical same period data and daily environmental data to complete missing values for subjective symptom data, using adjacent node data interpolation to complete missing values for environmental interaction data, and using prompt completion + default rule to complete missing values for static basic data; S13: performing standardization processing on the completed multi-modal data.

3. A method of assessing the risk of an acute asthma exacerbation in a child according to claim 2, wherein, The dynamic weight feature extraction model in step S2 comprises a basic feature layer, a dynamic weight layer and a fusion feature layer, and the specific process of extracting specified features is as follows: S21: extracting specified basic features through the basic feature layer, including the target child's acute attack frequency in a specified time period, PEF diurnal variation rate, AQI exceeding cumulative duration, and allergen daily exposure intensity, to obtain a basic feature vector; S22: the dynamic weight layer gives different specified basic features dynamic differentiated weights through an attention mechanism, multiplies the basic features by the corresponding dynamic differentiated weights to obtain a dynamic weight vector: S23: multiplying the basic feature vector and the dynamic weight vector to generate a weighted feature vector; S24: incorporating the child's medication compliance in a specified time period as a risk adjustment factor to adjust the numerical value of the weighted feature vector, reflecting the risk mitigation effect of regular medication or the risk amplification effect of irregular medication, and constructing a fusion feature vector.

4. The method of claim 3, wherein the method further comprises: The specific process of inputting the fusion feature vector into the logistic regression model in step S3 to output the probability of asthma acute attack occurring in the future specified period is as follows: S31: Load the pre-trained model parameters, including the feature coefficients β0, β1, β2, β3, and β4, where β0 is the intercept term, and β1-β4 correspond to the four dimensions of the fusion feature vector; S32: Input the fusion feature vector into the logistic regression model for linear combination calculation to obtain the linear prediction value z reflecting the risk level: S33: Convert the linear value to probability through Sigmoid function mapping.

5. The method of claim 1, wherein the method further comprises: The specific process of step S4 is as follows: S41: Load the XGBoost model parameters, including tree structure parameters, feature attention coefficients, and base probability mapping parameters; S42: Based on the loaded feature attention coefficients, assign initial importance weights to each dimension of the fusion feature vector; S43: The XGBoost model calculates through multiple parallel decision trees, integrates the results, and outputs the original risk score; S44: Convert the original risk score to the attack probability through an improved Sigmoid function; S45: Map the probability to the preliminary risk level.

6. The method of claim 4, wherein the method further comprises: The specific process of step S5 is as follows: S51: Obtain the target child's acute attack records in the past year to construct the target child's historical attack feature vector; S52: Calculate the cosine similarity between the current fusion feature vector and H1, H2, and H3, respectively; S53: Take the average of multiple historical attack similarities as the final calibration basis; S54: Calibrate the preliminary risk level based on the average similarity and the preset calibration rules.

7. The method of claim 1, wherein the method further comprises: It also includes setting a feedback adjustment mechanism, the specific process is as follows: S7: Obtain the feedback data of the target child, including intervention execution data and disease change data; S8: Establish a mapping relationship between the feedback data and the risk assessment results, evaluate the accuracy of the risk assessment results according to the mapping relationship, and quantify the deviation between the risk assessment results and the feedback data; S9: Based on the deviation between the risk assessment results and the feedback data, optimize the dynamic weight feature extraction model, the logistic regression model, and the XGBoost model, and re-execute steps S2-S6.

8. A system for assessing the risk of an acute asthma exacerbation in a child for implementing a method for assessing the risk of an acute asthma exacerbation in a child according to any one of claims 1 to 7, characterized in that, It includes: Multi-modal data acquisition and preprocessing module: used to acquire the target child's static basic data, dynamic physiological data, environmental interaction data, and subjective symptom data, and preprocess these multi-modal data; Dynamic weight feature extraction model, used to extract specified features to construct a basic feature vector, adjust dynamic feature weights based on individual differences and real-time scenarios, construct a weighted feature vector based on the basic features and dynamic feature weights, and incorporate the child's recent medication compliance to construct a fusion feature vector; Logistic regression model: used to input the fusion feature vector, output the probability of asthma acute attack occurring in the future specified period, send a push regular care plan instruction to the scheme pushing module when the probability is less than the first probability threshold, send a push emergency intervention scheme instruction to the scheme pushing module when the probability is greater than the second probability threshold, or send an evaluation instruction to the risk level assessment module; XGBoost model: After receiving the evaluation instruction, input the fusion feature vector into the XGBoost model, output the preliminary risk level, and the preliminary risk level includes low risk, medium risk and high risk; Risk level calibration module: Calculate the cosine similarity between the current fusion feature vector and the historical onset feature vector, and calibrate the preliminary risk level based on the cosine similarity to obtain the calibrated risk level; Dynamic intervention scheme generation and pushing module: Based on the calibrated risk level, generate a dynamic intervention scheme. When the calibrated risk level is low risk, generate the corresponding routine care scheme; when the risk level is medium risk, generate the intensive monitoring scheme; and when the risk level is high risk, output the corresponding emergency intervention scheme.

Images