A new-born low blood sugar early warning system based on data fusion

Abstract

The present application belongs to the technical field of clinical medicine, and aims at the deficiency of the existing new-born low blood sugar early warning technology, and provides a new-born low blood sugar early warning system based on data fusion, a model is established based on genetic algorithm, the influence of key features on blood sugar is determined through multi-objective fitness function design and a series of operations, and early warning information is generated.The present application fuses multi-source data, optimizes model training, balances multi-objectives, and improves the accuracy and reliability of new-born low blood sugar early warning.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of clinical medicine and relates to a new-born low blood sugar early warning system based on data fusion. BACKGROUND

[0002] New-born low blood sugar is a common metabolic disorder disease in the new-born period, and if not discovered and treated in time, it may cause irreversible damage to the nervous system development of the new-born.

[0003] In the prior art, only the blood sugar of the new-born or a small amount of vital signs are relied on, the influence of the mother factor on the blood sugar of the new-born is ignored, and the early warning is not comprehensive enough. And the data from different sources cannot be effectively integrated to extract useful information.

[0004] And in the prior art, linear regression or simple machine learning algorithms are mostly used, and multiple focuses are laid on a single target, and the balance of multiple targets such as sensitivity and specificity is ignored, which may lead to missed high-risk cases. SUMMARY

[0005] To solve the problems in the background art, the application provides a new-born low blood sugar early warning system based on data fusion.

[0006] To achieve the above purpose, the technical scheme adopted by the application is as follows:

[0007] A new-born low blood sugar early warning system based on data fusion comprises:

[0008] The data processing module establishes a genetic algorithm model;

[0009] The genetic algorithm model is used to analyze the fusion data related to the new-born low blood sugar, and the key feature data affecting the blood sugar is determined;

[0010] Based on the judgment of the key feature data affecting the blood sugar, low blood sugar early warning information is generated;

[0011] The processing optimization module evaluates and optimizes the genetic algorithm model based on the accuracy of the low blood sugar early warning information.

[0012] Further, the fusion data related to the new-born low blood sugar is obtained by the data acquisition module; the fusion data comprises detection data of the mother's milk and blood samples collected at a preset time point after delivery, and detection data of the new-born vital signs.

[0013] Further, the method for constructing the data set of the fusion data is as follows:

[0014] The correlation analysis method is used to screen out features closely related to the blood glucose of the newborn from the mother's emulsion sample data, the mother's blood sample data and the vital sign data of the newborn, remove redundant features and reduce the data dimension; the feature is a detection item of the detection data;

[0015] The screened features are spliced according to the set order to construct a fusion feature vector;

[0016] A time decay factor is designed for the mother's emulsion sample data at different time points after childbirth;

[0017] During feature splicing, attention weights are introduced through a genetic algorithm to dynamically amplify the influence of the top L features with high correlation, where L is a positive integer.

[0018] Further, the specific method for the data processing module to establish a genetic algorithm model is:

[0019] Based on multi-objective, an NSGA-II algorithm is used to generate a Pareto optimal solution set through non-dominated sorting and crowding comparison;

[0020] Real number coding is used to code the model parameters;

[0021] A certain number of initial individuals are randomly generated to form an initial population;

[0022] Each feature obtained by the data acquisition module is set as an individual to generate a population.

[0023] Further, the fitness function design based on multi-objective includes:

[0024] Objective 1: Minimize blood glucose prediction error through MSE;

[0025] Objective 2: Minimize model complexity through L1 regularization term;

[0026] Objective 3: Set the recall rate to maximize the early warning sensitivity.

[0027] Further, the specific method for the data processing module to analyze the data set containing multi-source features based on the genetic algorithm model to determine the proportion of the influence of key feature data on blood glucose is:

[0028] The genetic algorithm model selects the top M individuals with high fitness values from the current population using a tournament selection method, where M is a positive integer, as the parents of the next generation population;

[0029] The selected parent individuals are subjected to a crossover operation to generate new child individuals by exchanging the data content represented by each parent individual;

[0030] The cross probability is initially set as a first value, and gradually decreases with the increase of the number of iterations;

[0031] The mutation probability is initially set as a second value, and is automatically increased when the population fitness variance is lower than a threshold value;

[0032] The top 10% individuals with the highest fitness in each generation are directly entered into the next generation;

[0033] The global search is performed by using the genetic algorithm, and then the local fine tuning is performed;

[0034] When the fitness improvement of a preset generation is less than 1% continuously, the local search stage is automatically switched to;

[0035] Based on the trained genetic algorithm model, the feature importance evaluation method is used to calculate the contribution degree of each individual to the blood glucose prediction result, i.e., the feature importance score, and the higher the feature importance score is, the greater the influence of the feature on blood glucose is.

[0036] Further, the specific method for performing the global search by using the genetic algorithm, quickly approaching the optimal solution region, and then performing the local fine tuning is as follows:

[0037] The appearance frequency and / or corresponding value of any individual are corrected, the correction of the corresponding value of the individual is the number of times of changing in a set range,

[0038] The combination relationship between each individual in the population and different individuals is recorded;

[0039] The influence of the change of the current individual on the blood glucose prediction result is judged, and the influence function of the correction of the current individual on the blood glucose prediction result is simulated by using a polynomial regression equation.

[0040] Further, the processing optimization module evaluates the genetic algorithm model by using evaluation indexes including the precision, recall rate and F1 value, and optimizes the genetic algorithm model according to the evaluation result of the genetic algorithm model.

[0041] Further, a distributed computing architecture is used, the Spark or Dask framework is used to parallelize the fitness calculation of the genetic algorithm, the population individuals are distributed to different computing nodes for evaluation, and the matrix operation in the mutation stage is accelerated by using CUDA.

[0042] Compared with the prior art, the present application has the following beneficial effects:

[0043] The present application fuses multi-source data of mother emulsion components, mother blood indexes and newborn vital signs, constructs a more comprehensive feature system, and improves the capturing ability of the model on complex metabolic correlation.

[0044] The application drives population evolution through fitness function, dynamically adjusts the influence proportion of features on blood glucose, and combines adaptive crossover / mutation probability and optimization strategy to improve model convergence speed and accuracy.

[0045] The application simultaneously optimizes prediction error, model complexity and recall rate to generate a Pareto optimal solution set. BRIEF DESCRIPTION OF DRAWINGS

[0046] Fig. 1 is a system structure diagram of the application;

[0047] Fig. 2 is a model training principle diagram of the application;

[0048] Fig. 3 is a newborn early warning principle diagram of the application. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0050] As shown in Figs. 1-3 , the technical solutions adopted by the application are as follows: a newborn hypoglycemia early warning system based on data fusion, comprising:

[0051] A data acquisition module acquires mother's milk sample and mother's blood sample at a preset time point after delivery, detects components in a preset item list, and obtains newborn vital sign data.

[0052] A data preprocessing module performs data cleaning and data standardization processing on the acquired mother's milk sample data, mother's blood sample data and newborn vital sign data.

[0053] A data fusion module combines the preprocessed mother's milk sample data, mother's blood sample data and newborn vital sign data to form a data set containing multiple source features.

[0054] A data processing module establishes a genetic algorithm model, analyzes the data set containing multiple source features based on the genetic algorithm model, determines the influence proportion of multiple data on blood glucose, and selects key feature data.

[0055] A judgment module generates hypoglycemia early warning information based on key feature data judgment.

[0056] The processing optimization module evaluates and optimizes the genetic algorithm model based on the accuracy of the hypoglycemia early warning information.

[0057] The present application aims to early warn neonatal hypoglycemia through multi-source data fusion and genetic algorithm model. The system mainly consists of data acquisition, preprocessing, fusion, processing, judgment and processing optimization function modules, and each function module works cooperatively to improve the accuracy and reliability of early warning.

[0058] Systematic sampling is performed at five time points of 1h, 3h, 6h, 12h and 24h after the birth of the newborn, and the sampling content includes neonatal vital sign data, and mother's milk sample and mother's blood sample. The neonatal vital sign data includes heart rate, respiratory rate, body temperature, blood oxygen saturation and other indicators, the mother's milk sample includes lactose, fat, immunoglobulin and other indicators, and the mother's blood sample includes blood glucose, insulin, inflammatory factor and other indicators. The components in the preset project list are detected, and the neonatal vital sign data is obtained. Multi-source data acquisition can comprehensively obtain information related to neonatal blood glucose. The components in the mother's milk and blood samples may affect the nutrient intake and metabolism of the newborn, and the neonatal vital sign data directly reflects the physical condition of the newborn. Integrating these data can provide more abundant and accurate basis for subsequent analysis and early warning.

[0059] The collected data is subjected to data cleaning and data standardization processing, and heterogeneous data such as mg / dL blood glucose value and bpm heart rate are normalized, interpolation method or rejection of incomplete samples is used, and abnormal data is identified and corrected using box plot. Data cleaning can remove noise, errors and missing values in the data, improving the quality of the data. Data standardization processing makes data from different sources and different scales comparable, which helps subsequent model training and analysis, and improves the accuracy and stability of the model.

[0060] After obtaining the mother's milk sample data, mother's blood sample data and neonatal vital sign data, the data fusion module is used for data fusion. The specific method of data fusion is as follows:

[0061] Correlation analysis method is used to screen out the features in the mother's milk sample data, mother's blood sample data and neonatal vital sign data that are closely related to the neonatal blood glucose, including lactose concentration, fat content, protein content, free fatty acid, insulin, leptin, cortisol, fasting blood glucose, postprandial blood glucose, gestational age and neonatal body temperature. These data are removed from redundant features to reduce the data dimension.

[0062] The correlation analysis method is used to screen features closely related to the blood glucose of the newborn, wherein the features are detection items in detection data including lactose concentration, fat content, protein content, free fatty acid, insulin, leptin, cortisol, fasting blood glucose, postprandial blood glucose, gestational age and newborn body temperature, etc. The feature data is removed of redundant features to reduce the data dimension. The fused features include mother baseline data such as average blood glucose during pregnancy, and newborn real-time vital signs such as heart rate change rate. The features are time-aligned and the mother-infant data are time-stamped matched, such as the association of the mother's postprandial 2-hour data with the corresponding period of the newborn's signs. The correlation analysis screening features and reducing data dimension can reduce the complexity of the data, improve the calculation efficiency, and at the same time avoid the interference of redundant information on the model.

[0063] The screened features are spliced in the set order to construct a fused feature vector. The construction of the fused feature vector integrates the multi-source data into a unified data set, which is convenient for subsequent model processing.

[0064] A time decay factor is designed for the mother's milk sample data at different time points after childbirth. The time decay factor takes into account the influence of the mother's postpartum time on the composition of the milk, making the data more in line with the actual situation.

[0065] In the feature splicing, attention weights are introduced through a genetic algorithm to dynamically amplify the influence of the top N features with high correlation, wherein N is a positive integer and can be set by the user. The introduction of attention weights can highlight the features with high correlation with the blood glucose of the newborn, enhance the model's ability to capture key information, and improve the accuracy of the early warning.

[0066] Based on the above method, preprocessed data is collected, and the model is trained using these data.

[0067] First, a genetic algorithm model is established, and a fitness function is designed based on multiple objectives, including:

[0068] Objective 1: Minimize blood glucose prediction error through MSE.

[0069] Objective 2: Minimize model complexity through L1 regularization term.

[0070] Objective 3: Set the recall rate to maximize the sensitivity of the early warning.

[0071] The NSGA-II algorithm is used to generate a Pareto optimal solution set through non-dominated sorting and crowding comparison. Real number coding is used to encode the model parameters. Each feature is set as an individual to generate a population. A certain number of initial individuals are randomly generated to form an initial population. Real number coding and random generation of the initial population provide a suitable search space and starting point for the genetic algorithm.

[0072] The multi-objective fitness function design comprehensively considers blood glucose prediction error, model complexity and early warning sensitivity, so that the model achieves a good balance in multiple aspects, improving the performance and practicality of the model.

[0073] The genetic algorithm model extracts the influence level of different data on neonatal blood glucose.

[0074] The genetic algorithm model selects the top M individuals with high fitness from the current population as the parents of the next generation according to the fitness value of each individual, where M is a positive integer.

[0075] The selected parent individuals are subjected to crossover operation to generate new child individuals by exchanging the data content represented by each parent individual.

[0076] The crossover probability is initially set to a first value and gradually decreases as the number of iterations increases; the mutation probability is initially set to a second value and automatically increases when the population fitness variance is below a threshold.

[0077] Tournament selection, crossover and mutation operations are the core of genetic algorithms, which constantly evolve the population to find the optimal solution. Dynamic adjustment of crossover and mutation probabilities can balance global and local search capabilities and avoid falling into local optima.

[0078] The top 10% individuals with the highest fitness are directly entered into the next generation. Retaining the top individuals can ensure that the excellent genes of the population are passed on and accelerate the convergence of the algorithm.

[0079] Global search through genetic algorithms quickly approaches the optimal solution region, followed by local fine-tuning. The combination of global search and local fine-tuning can quickly find the approximate range of the optimal solution, then make precise adjustments in the local area, improving search efficiency and accuracy.

[0080] When the continuous preset generation fitness improvement is less than 1%, automatically switch to the local search stage.

[0081] Based on the trained genetic algorithm model, the feature importance evaluation method is used to calculate the contribution of each fused feature to the blood glucose prediction result, i.e. feature importance score. Feature importance evaluation can clearly determine the influence of each data feature on blood glucose prediction, providing more targeted basis for subsequent early warning judgment.

[0082] The genetic algorithm model includes global search and local fine-tuning, which records the appearance frequency or change frequency of each individual and the combination relationship between individuals in the population.

[0083] The influence of the current individual's change on the blood glucose prediction result is judged, and the influence function of the current individual frequency on the blood glucose prediction result is simulated by a polynomial regression equation.

[0084] The local optimal solution of the genetic algorithm model can be avoided by global search and local fine tuning. When some hormone level data of the mother in the training data is limited by the need for special reagents and laboratory equipment for hormone detection, the cost is high, the sample size is small, but the influence on blood glucose is large.

[0085] The appearance frequency and corresponding value of the individual represented by such features are corrected, and the correction of the individual corresponding value is the number of changes with a set amplitude. The combination relationship between each individual in the population and different individuals is recorded, and the specific formula is:

[0086]

[0087] Among them is the appearance frequency of the individual in the current generation, that is, the probability of being selected as the parent, which reflects the activity of the individual in the population. The higher the frequency, the more likely the genetic information is to be preserved; In the first generation, the probability of being selected as the parent reflects the activity of the individual in the population. The higher the frequency, the more likely the genetic information is to be preserved; is the learning rate, that is, the step size parameter of gradient descent, which controls the amplitude of frequency adjustment. By adjusting the learning rate update speed, it can avoid oscillation or non-convergence caused by too large step size. is the partial derivative of the blood glucose prediction error with respect to the appearance frequency, which reflects the sensitivity of the frequency adjustment to the error. A positive value indicates that increasing the frequency will increase the error, and the frequency needs to be reduced. A negative value is the opposite; is the linear term coefficient of in the polynomial regression, which represents the direct influence of the appearance frequency on the blood glucose prediction error; is the quadratic term coefficient of in the polynomial regression, which represents the nonlinear influence of the frequency; is the interaction term coefficient of and in the polynomial regression, which represents the synergistic effect of the frequency and the mutation probability; is the mutation probability of the individual .

[0088] Such individuals may not be inherited in the genetic algorithm model, and may be ignored due to the characteristics of the genetic algorithm model. Adding the appearance frequency and corresponding value changes of such feature data and simulating them through a polynomial regression equation can predict the influence of different feature data on blood glucose and avoid the genetic algorithm model from falling into an optimal solution.

[0089] Based on the key feature data output by the genetic algorithm model, low blood sugar early warning information is generated. According to the proportion of data on the influence of blood sugar, the risk of neonatal hypoglycemia can be more accurately identified, and early warning information can be timely issued, which helps medical staff to take corresponding measures to protect the health of newborns.

[0090] Meanwhile, the genetic algorithm model can also be optimized based on the accuracy of the low blood sugar early warning information through the processing optimization module. The genetic algorithm model is evaluated by using evaluation indexes including precision, recall rate and F1 value, and the genetic algorithm model is optimized according to the evaluation result of the genetic algorithm model. The genetic algorithm model is evaluated by using scientific evaluation indexes, so that the performance of the model can be comprehensively understood, and the problems existing in the genetic algorithm model can be found. According to the evaluation result, the accuracy and reliability of the genetic algorithm model can be continuously improved, so that the early warning system is more perfect.

[0091] In addition, the present application adopts a distributed computing architecture, uses the Spark or Dask framework to parallelize the fitness calculation of the genetic algorithm, and distributes the population individuals to different computing nodes for evaluation. The matrix operation in the mutation stage is accelerated by using CUDA. Distributed computing and parallel processing can fully utilize the resources of multiple computing nodes, improve the computing efficiency, and shorten the model training time. CUDA acceleration can further speed up the matrix operation and improve the overall performance of the system, so that the system can process a large amount of data more quickly and issue warnings in time.

[0092] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features, as long as the modifications, equivalent replacements or improvements are within the spirit and principles of the present application. Any modification, equivalent replacement, improvement, etc. shall be included in the protection scope of the present application.

Claims

1. A data fusion based neonatal hypoglycemia warning system, characterized in that, The application relates to a method for generating a low blood sugar early warning information of a newborn, comprising the following steps: a data processing module establishes a genetic algorithm model; based on the genetic algorithm model, fusion data related to the low blood sugar of the newborn is analyzed to determine key characteristic data affecting blood sugar; based on the judgment of the key characteristic data affecting blood sugar, the low blood sugar early warning information is generated; a processing optimization module evaluates and optimizes the genetic algorithm model based on the accuracy of the low blood sugar early warning information; the fusion data related to the low blood sugar of the newborn is obtained by a data acquisition module; the fusion data comprises detection data of mother's milk and blood samples collected at a preset time point after delivery and detection data of the vital signs of the newborn; a method for constructing a data set of the fusion data is as follows: correlation analysis is used to screen out characteristics closely related to the blood sugar of the newborn from the mother's milk sample data, the mother's blood sample data and the vital sign data of the newborn, remove redundant characteristics and reduce the data dimension; the characteristics are detection items of the detection data; the screened characteristics are spliced in a set order to construct a fusion characteristic vector; a time decay factor is designed for the mother's milk sample data at different time points after delivery of the mother; attention weights are introduced by using the genetic algorithm during the characteristic splicing, and the influence of the first L characteristics with high correlation is dynamically amplified, wherein L is a positive integer.

2. The data fusion based neonatal hypoglycemia warning system according to claim 1, wherein, the specific method for the data processing module to establish the genetic algorithm model is as follows: a multi-objective fitness function is designed, NSGA-II algorithm is adopted, non-dominated sorting and congestion degree comparison are performed, and a Pareto optimal solution set is generated; real number coding is adopted to code the model parameters; a certain number of initial individuals are randomly generated to form an initial population; each characteristic obtained by the data acquisition module is set as an individual to generate the population.

3. The data fusion based neonatal hypoglycemia warning system according to claim 2, wherein, the multi-objective fitness function design comprises the following steps: target 1: minimizing blood sugar prediction error by using MSE; target 2: minimizing model complexity by using an L1 regularization term; target 3: maximizing early warning sensitivity by setting recall rate.

4. The data fusion based neonatal hypoglycemia warning system according to claim 2, wherein, the specific method for the data processing module to analyze the data set containing multi-source characteristics based on the genetic algorithm model to determine the proportion of key characteristic data affecting blood sugar is as follows: the genetic algorithm model selects the first M individuals with high fitness from the current population as the parents of the next generation population by using the tournament selection method according to the fitness values of the individuals, wherein M is a positive integer; the selected parent individuals are subjected to a crossover operation to generate new child individuals by exchanging the data contents represented by each parent individual; the crossover probability is initially set as a first value and gradually decreases with the increase of the iteration number; the mutation probability is initially set as a second value and is automatically increased when the population fitness variance is lower than a threshold value; 10% of the individuals with the highest fitness are directly entered into the next generation; the global search is performed by using the genetic algorithm to quickly approach the optimal solution region, and then local fine tuning is performed; when the fitness is continuously improved by less than 1% for a preset number of generations, the local search stage is automatically switched to. Based on the trained genetic algorithm model, the feature importance evaluation method is used to calculate the contribution of each individual to the blood glucose prediction result, that is, the feature importance score. The higher the feature importance score, the greater the influence of the feature on blood glucose.

5. The data fusion based neonatal hypoglycemia warning system according to claim 4, wherein, The specific method is to quickly approach the optimal solution region through global search of genetic algorithm, and then perform local fine tuning: Correcting the appearance frequency and / or corresponding value of any individual, the correction of the individual corresponding value is the number of changes with a set amplitude, Record the combination relationship between each individual in the population and different individuals; Judge the influence of the current individual change on the blood glucose prediction result, and simulate the influence function of the current individual modification on the blood glucose prediction result through a polynomial regression equation.

6. The data fusion based neonatal hypoglycemia warning system according to claim 1, wherein, The processing optimization module evaluates the genetic algorithm model using evaluation indicators including precision, recall and F1 value, and optimizes the genetic algorithm model according to the evaluation result of the genetic algorithm model.

7. The data fusion based neonatal hypoglycemia warning system according to claim 1, wherein, Using distributed computing architecture, using Spark or Dask framework to parallelize the fitness calculation of genetic algorithm, and distributing population individuals to different computing nodes for evaluation; using CUDA to accelerate matrix operation in the mutation stage.

Images