CN116030972A - Health evaluation system and method based on multi-layer perceptron neural network model - Google Patents

Abstract

The invention relates to a health evaluation system and method based on a multi-layer perceptron neural network model, which comprises the steps of firstly extracting samples from a medical database, dividing the number of training samples and the number of testing samples according to training test proportion, constructing a simulated neural network structure model, readjusting the health status and influencing factors in the health evaluation process of different crowds by adopting a decision technology, a prediction technology and a simulation technology to control a management mode in real time, carrying out iterative verification on the simulated neural network structure model by using the test sample, improving the accuracy of the evaluation mode by optimizing the smoothing factors of a network in the iterative verification process, and continuously optimizing the whole model evaluation effect after obtaining a final model. Based on the conventional evaluation of the health condition of the crowd, the health state of the crowd is visually evaluated by using the multi-layer perceptron neural network model in the data mining technology, so that a novel health evaluation mode is constructed, and the health risk management and control measure selection can be assisted.

Description

Health evaluation system and method based on multi-layer perceptron neural network model

[ field of technology ]

The invention relates to the technical field of human health assessment, in particular to a health assessment system and method based on a multi-layer perceptron neural network model.

[ background Art ]

Generally, health assessment can be subdivided into narrow and broad terms. The narrowly defined health assessment is that a doctor comprehensively judges the health condition of a patient by inquiring the physical and psychological states of the patient before diagnosing the patient. The generalized health assessment focuses on the past medical history, family genetic history, family medication and other conditions of patients, and focuses on the health status, family economic level, social environment and the like of other family members. Meanwhile, along with the continuous enrichment of health connotation, the category of health assessment is gradually expanded, and the simple health assessment provided by the intelligent algorithm technology and the big data technology is gradually one of ways for people to pay attention to self health.

The existing health assessment and management methods put more attention on symptom diagnosis, scale health assessment and quantitative diagnosis, but the methods depend on professional medical background and knowledge, and lack a universal research method for the system. The existing research on health assessment is mainly focused on aspects of brain electrical signals, electrocardiosignals, DNA related to diseases and the like, and human health assessment based on physiological parameters is rare. In order to further improve the accuracy and stability of human health assessment, scholars at home and abroad propose a plurality of intelligent algorithms such as genetic algorithm, lifting method, fuzzy set and the like. These algorithms, while somewhat successful, have limitations. Therefore, how to simultaneously improve the high performance of health evaluation and control evaluation of people becomes a problem of urgent need for research. There is no suitable algorithm or system to solve the above-mentioned problems.

Chinese patent application: CN114626469a, bulletin day: 2022.06.14A cerebral apoplexy limb rehabilitation method, device and system based on a multilayer perceptron model are disclosed, and the method comprises the following steps: prompting a user to perform motor imagery of limb actions; acquiring motor imagery electroencephalogram signals of a user; processing the motor imagery electroencephalogram signals to obtain sliding window electroencephalogram signal data; inputting the sliding window electroencephalogram data into a trained time-space domain multi-layer perceptron neural network model; obtaining a motor imagery identification result output by the time-space domain multi-layer perceptron neural network model; and guiding the neuromuscular electro-stimulator to perform electro-stimulation treatment on the muscles of the limbs of the user according to the motor imagery identification result. The cerebral apoplexy limb rehabilitation method provided by the invention can realize accurate acquisition, effective identification and correct classification of EEG signals, and can help a user to finish rehabilitation training by identifying the action intention of the user and carrying out electric stimulation treatment on hand muscle nerves.

Although the above patent provides a method, a device and a system for rehabilitation of cerebral apoplexy limb based on a multi-layer sensor model, the method cannot be used in the field of health evaluation, and a system and a method for health evaluation based on a multi-layer sensor neural network model as described in the application are not seen at present.

[ invention ]

The invention aims to provide a health evaluation system based on a multi-layer perceptron neural network model.

It is still another object of the present invention to provide a method for health assessment based on a multi-layer perceptron neural network model.

In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:

the health evaluation system based on the multi-layer perceptron neural network model is characterized by comprising a medical database and a multi-layer perceptron neural network algorithm model, wherein the medical database is used for health evaluation, and the multi-layer perceptron neural network algorithm model is used for operation of a complex database.

As a preferred example, the medical database includes a sample set and basic information and medical indexes of health data of different people, the sample set includes general demographic characteristics, main disease conditions, physiological parameters of human body and other parameters, and the basic information includes: age, sex, height, weight, BMI, etc.; the medical index includes: physiological parameters, biochemical parameters, blood lipid index, electrocardiogram frequency, etc.

As another preferred example, the multi-layer perceptron network neural algorithm model includes the following steps: s01, determining the number of hidden layers; s02, determining the number of hidden layer nodes; s03, determining a learning rate factor alpha; s04, selecting a momentum factor beta; s05, selecting an activation function; s06, determining a training test proportion; s07, determining an optimization algorithm.

More preferably, the main conditions for calculating the number of explicitly hidden layer nodes in the step S02 are as follows: the sum of the total node number must be less than N-1, N being the number of training samples; the connection weight of the network structure model is smaller than the number of training samples, and the multiple between the connection weight and the training samples is 2-10 times.

More preferably, in the iterative process, the new learning rate is the original learning rate multiplied by a constant Φ, and the expression is: α (t+1) =α (t) ×Φ, and when the total error function VE < 0, the value range of the constant Φ is: phi is more than 0.001 and less than 1; when the total error function VE > 0, the value range of the constant phi is as follows: phi is more than 1 and less than 50.

More preferably, in the adjusting and modifying process of the weight, the step S04 uses the previous weight change to smooth the learning path, where the expression is: w (t+1) =w (t) +βvw (t), wherein W (t) is a main weight in the t-th iteration, βvw (t) is a direction of adjustment change of the memory t weight, and a main range of values of the constant β is 0-t.

More preferably, the Sigmoid function expression adopted in the step S05 is:

the parameter A is used for adjusting the overall change trend and range of the data, and promoting the change rate of the original gentle region to be continuously accelerated, and the value range is as follows: a is more than 0.01 and less than 100; the parameter B is generally 1, so that the function sensitivity is higher; the parameter C is used for eliminating an insensitive interval in the data, continuously increasing a change area of the data, and the value range is as follows: c is more than 1 and less than 100.

In order to achieve the second purpose, the invention adopts the following technical scheme:

a health assessment method based on a multi-layer perceptron neural network model is characterized by comprising the following steps:

s1, constructing a simulated neural network structure model, which comprises

S11, respectively dividing the training sample number and the test sample number according to the training test proportion;

s12, performing simulation calculation on relations between diseases and influence factors in different crowds by using training samples;

s13, learning and training among a limited number of samples, and eliminating the correlation and complexity among variables;

s2, training a sample iteration model, and preliminarily obtaining a final model, wherein the step comprises the following steps:

s21, readjusting the health status and influencing factors in the health evaluation process of different people by adopting a decision-making technology, a prediction technology and a simulation technology, and controlling the management mode in real time;

s22, feeding back the human health data information under various conditions in real time;

s23, carrying out iterative verification on the simulated neural network structure model by using a test sample;

s24, after reaching the meaning rule of the satisfactory approximation integral sample, preliminarily obtaining a final model;

s3, optimizing the overall model evaluation effect, including:

s31, optimizing a smoothing factor of a network to improve the accuracy of an evaluation mode in the process of training a sample iteration model;

s32, performing simulation experiment verification comparison, and reducing the number of neurons in the hidden layer;

s33, complexity of the model is reduced, and the overall model evaluation effect is optimized.

As a preferred example, the number of training samples and the number of test samples are obtained from a medical database.

The invention has the advantages that:

1. on the basis of conventional evaluation of the health condition of the crowd, the method crosses, fuses and permeates modern medical treatment and engineering thinking, and adopts the multi-layer perceptron neural network model in the data mining technology to visually evaluate the health condition of the crowd, so that the evaluation result is more convincing by multi-layer and multi-dimensional expansion, a novel health evaluation mode is constructed, and health risk management and control measures can be assisted to select.

2. According to the invention, the physiological parameters and other parameters of the human body are formed into a sample set for training, the output results at the same moment are classified into healthy, sub-healthy and unhealthy through parameter change and interaction conditions, and finally, different classification results are monitored to evaluate the health condition of the human body.

3. According to the invention, the influence factors of the diseases are analyzed in combination with the disease changes, the development trend of the diseases of the crowd is evaluated, weak links of disease prevention and control are identified, a universal and high-accuracy remote health evaluation method is found, and scientific reference basis can be provided for epidemiological research and dynamic development rules of the diseases of the crowd, hazard research and formulation of prevention and control strategies.

[ description of the drawings ]

FIG. 1 is a flow chart of a clinical practice of a health assessment system and method based on a multi-layer perceptron neural network model of the present invention.

FIG. 2 is a view showing the observation prediction of example 2.

Fig. 3 is a comparison chart of the importance ranking of input layer parameters of example 2.

FIG. 4 is a graph of the model ROC of example 2.

FIG. 5 is a view showing the observation prediction of example 3.

Fig. 6 is a comparison chart of the importance ranking of input layer parameters of example 3.

FIG. 7 is a graph of the model ROC of example 3.

FIG. 8 is a view showing the observation prediction of example 4.

Fig. 9 is a comparison chart of the importance ranking of input layer parameters of example 4.

FIG. 10 is a graph of the model ROC of example 4.

FIG. 11 is a diagram of a health assessment model of a system and method for health assessment based on a multi-layer perceptron neural network model of the present invention.

[ detailed description ] of the invention

The invention is further described below with reference to examples and with reference to the accompanying drawings.

Example 1

1. Material method

Referring to fig. 1, fig. 1 is a flowchart of a clinical practice of a health assessment system and method based on a multi-layer perceptron neural network model of the present invention, comprising:

1. establishing a medical database for health assessment

Based on the physical examination data of a certain three-hospital in the past year, after processing such as integration, coding, checking and presenting repeated values, checking and checking are performed by using R software, and the error, repeated or questionable data is checked again, comprising the following steps: removing missing data, such as partial physical examination data missing age and recording information of blood fat index results; excluding duplicate data; excluding abnormal data, such as physical examination data with body mass index of unit number; removing logic error data; and finally, checking the collected data with the original data one by one to ensure the accuracy and reliability of the data.

The finally established medical database of the three hospitals comprises a sample set and basic information and medical indexes of health data of different crowds, wherein the sample set comprises general demographic characteristics, main disease conditions, physiological parameters of human bodies and other parameters, and the basic information comprises: age, sex, height, weight, BMI, etc.; the medical index includes: physiological parameters, biochemical parameters, blood lipid index, electrocardiogram frequency, etc.

2. Establishing a multi-layer perceptron neural network algorithm model for complex database operation

2.1 determination of the number of hidden layers

Generally, the number of layers of the hidden layer is increased, so that errors among network structures can be reduced, and the measuring and calculating precision is improved. Meanwhile, the network structure is complicated, the training time of the network structure is increased, and the phenomenon of over fitting occurs. Before designing a multi-layer perceptron neural network (MLPNN), first, a 3-layer MLPNN should be considered, namely an input layer, an output layer, and a hidden layer, which may be one or more layers. If a linear function is used as the activation function of both the input layer and the output layer, a Sigmoid function is used as the activation function of the hidden layer, then an MLPNN with 1 hidden layer can be infinitely close to any rational function with arbitrary precision.

2.2 determination of hidden layer node count

The number of hidden layer nodes is a direct factor causing 'overfitting' in the training process, and has great influence on the performance of the network structure model. In order to ensure that the performance of the network structure is high enough, the generalization capability of the model is improved, the phenomenon of over fitting in the training process is avoided, and the basic principle of node number measurement is as follows: under the condition of meeting the precision requirement, the number of hidden nodes is as small as possible, and a compact network structure is maintained.

The main conditions for calculating the number of the nodes of the clear hidden layer are as follows: (1) the sum of all node numbers is required to be smaller than N-1, wherein N is the number of training samples, and the system error of the network model tends to zero irrespective of the characteristics of the training samples, namely the established network model has no generalization capability and has no practical value; (2) the connection weight of the network structure model is necessarily smaller than the number of training samples, the multiple between the connection weight and the number of the training samples is 2-10 times, the samples are divided into a plurality of parts, and the reliability of the network model is improved by combining the 'alternate training' method.

2.3 determination of learning Rate factor alpha

The learning rate factor alpha directly determines the overall learning speed and whether the overall learning can be achieved, and if the learning speed is too slow, the learning accuracy is possibly reduced, and meanwhile vibration is easy to occur. The initial value of the learning rate factor alpha is selected by only arbitrarily defining a positive number between 0 and 1. After the learning weight is set, training is started. Each learning rate change is determined by the change direction of the total error, and in the iterative process, the new learning rate is the original learning rate multiplied by a constant phi, so that the learning speed is increased, and the expression is: α (t+1) =α (t) ×Φ; the constant phi is manually valued, the convergence and the running speed of the whole training are directly affected by the constant phi, if the constant phi and the running speed are too large, local minima are easily generated, so that the convergence speed is affected, and under the general condition, when the total error function VE is smaller than 0, the constant phi is valued in the range of: phi is more than 0.001 and less than 1; when the total error function VE > 0, the value range of the constant phi is as follows: phi is more than 1 and less than 50.

2.4 selection of momentum factor beta

The momentum factor beta is used for preventing oscillation and improving convergence speed, and in the process of adjusting and modifying the weight, the last weight change is utilized to smooth the learning path, and the expression is as follows: w (t+1) =w (t) +βvw (t), wherein W (t) is a main weight in the t-th iteration, βvw (t) is a direction of adjustment change of the memory t weight, and a main range of values of the constant β is 0-t. If the value of beta is too large, the overall training speed can be increased to a certain extent, but convergence and precision are also deteriorated; if the value of β is too small, the modification and change of the weight of each round cannot be affected to a great extent, so that the specific value of β needs to be continuously adjusted and selected according to specific iteration conditions.

2.5 selection of an activation function

The adjusted Sigmoid function is as follows:

a, B, C is selected as three main parameters, wherein the parameter A is used for adjusting the overall change trend and range of data, promoting the change rate of the original gentle region to be continuously accelerated, and the value range is as follows: a is more than 0.01 and less than 100; the parameter B is generally 1, so that the function sensitivity is higher; the parameter C is used for eliminating an insensitive interval in the data, continuously increasing a change area of the data, and the value range is as follows: c is more than 1 and less than 100. By using the parameters A, B, C to adjust and modify overall data, the Sigmoid function has certain advantages in avoiding insensitive areas, but has defects in convergence speed training. The main parameter values of the activation function are thus properly selected.

2.6 determination of training test proportion

The ratio of training sample and test sample data is set to 7:3, training-checking-supporting 3 individual cases of the subareas are randomly distributed according to the relative number of the individual cases, the relative number of the training areas is 70%, the relative number of the checking areas is 30%, and the adherence area is 0%.

2.7 determination of optimization algorithms

Selecting a softMax function as an activation function of an output layer; and selecting the hyperbolic tangent function as a hidden layer activation function, and adopting an adjusted conjugate gradient algorithm for a neural network optimization algorithm. In view of the fact that the multi-layer perceptron neural network is a supervised learning algorithm, sensitivity, specificity, cut-off value, prediction accuracy, about log index, AUC, ROC curve and 95% credible interval are selected as evaluation standards, and a cyclic debugging method is adopted to determine optimal parameter values.

3. And screening chronic disease high incidence seeds in the physical examination data from a medical database, and evaluating the chronic disease incidence conditions of the crowd and influence factors thereof by combining a multi-layer perceptron neural network algorithm model.

4. And combining the predictive observation diagram and the importance ranking diagram, visually displaying the crowd health evaluation mode, wherein the abscissa is whether the crowd is ill, and the ordinate is the predictive pseudo probability. The multi-layer perceptron neural network model in the observation prediction graph defaults to 0.5 as a probability demarcation of whether the reservoir suffers from chronic disease to discriminate correctness and mistakes.

5. At present, most of researches on the disease conditions, influence factors and development trends of the main diseases of the crowd adopt logistic regression analysis, and the method is common and has general evaluation and prediction efficiency, so that the evaluation efficiency of the logistic regression model and the multi-layer perceptron neural network model are compared.

6. In the screening of the effective data samples of the longitudinal physical examination data 3462 group of nearly five years, the first three patients with high incidence of prostate hyperplasia, thyroid nodule and fatty liver were found, and the predicted efficacy level was evaluated by predicting the possibility of disease and influencing factors in the aged population, and displaying the predicted efficacy level in a visual manner in examples 2-4.

Example 2

The embodiment combines the prostate hyperplasia in the high-altitude diseases of the crowd to construct a crowd health assessment method based on a multi-layer perceptron neural network model, which comprises the following steps:

s1, a physical examination total 3462 groups of effective data samples, a model training sample is set to 2417 groups (69.8%), and a model test sample is set to 1045 groups (30.2%);

s2, standardizing input sample data of each layer according to different input variable dimensions in each layer of the input layer, the hidden layer and the output layer;

s3, setting contents of the dependent variables (input layers) as suffering from and not suffering from prostatic hyperplasia; the factor setting content is such variables as age, sex, BMI, blood pressure and the like; the covariates are quantitative data such as triglyceride, total cholesterol, low density lipoprotein cholesterol, high density lipoprotein cholesterol and the like; establishing a self-learning multi-layer perceptron neural network model;

and S4, drawing ROC curves of a binary logistic regression model and a multi-layer perceptron neural network model respectively by taking the cut-off value calculated by the predictive model formula as a check variable and taking whether the prostate hyperplasia exists as a state variable.

The test results are:

1) Specific model structure: the input layer is composed of 18 neurons including age, total cholesterol, sex, triglyceride, blood pressure, high density lipoprotein cholesterol, BMI, and low density lipoprotein cholesterol; the hidden layer is the deviation of each layer, and 8 neurons are all arranged; the output layer is composed of 2 neurons with and without prostatic hyperplasia. Details are given in table 1 below.

Table 1 neural network model structure of multilayer perceptron

2) And (3) observing a prediction graph: referring to fig. 2, fig. 2 is an observation prediction diagram of example 2. The established multi-layer perceptron neural network model is used for cross comparison of sample changes of each partition to form an observation prediction graph of the prostatic hyperplasia, and 0.5 is adopted as a correct and error probability demarcation for judging whether the reservoir is suffering from the prostatic hyperplasia by default. The ordinate of the predictive graph is the predictive pseudo-probability, and the abscissa is the cases with and without prostatic hyperplasia. According to the result of observing the prediction graph, the reservoirs are grouped according to the condition that the reservoirs are suffering from and not suffering from the prostatic hyperplasia, when any type of reservoir is not suffering from the prostatic hyperplasia as a prediction target, the prediction probability of the reservoir is obviously higher or lower than the possibility of suffering from other reservoirs, the difference between the reservoir and the reservoir is obvious, and the reservoir classification recognition effect is good.

3) Importance variable output: referring to fig. 3, fig. 3 is a comparison chart of importance ranking of input layer parameters according to embodiment 2. The conclusion output by the MLP is mainly influenced by the arguments of the input layer, thus ranking the importance of risk factors affecting prostate hyperplasia. The earlier single factor analysis results retained 8 risk factors affecting prostate hyperplasia. From the figure it is seen that gender, blood pressure, total cholesterol, low density lipoprotein cholesterol are important independent risk factors for prostate hyperplasia.

4) Evaluation effect:

(1) ROC curve

Referring to fig. 4, fig. 4 is a graph of the model ROC of example 2. Area under the model curve auc=0.753 of the multi-layer perceptron neural network, and area under the logistic regression model curve auc=0.610. The AUC of the multi-layer perceptron neural network model is greater.

(2) Evaluation index

The binary logistic regression model had 95% CI of 0.594-0.626, about dengue index maximum 0.1615, sensitivity of 49.54%, specificity of 66.61% and cutoff of 0.334, i.e., the probability of having prostatic hyperplasia was greater when the cutoff was greater than or equal to 0.334. The multilayer perceptron neural network evaluation model has 95% CI of 0.738-0.767, ROC curve about index maximum value of 0.4339, sensitivity of 45.57%, specificity of 97.81%, and cut-off value of 0.677, namely when the cut-off value is more than or equal to 0.677, the probability of suffering from prostatic hyperplasia is high. See table 2 below for details.

Table 2 comparison of the evaluation ability of the two models

Example 3

The embodiment combines thyroid nodules in high-incidence diseases of people to construct a crowd health assessment method based on a multi-layer perceptron neural network model, comprising the following steps:

s1, a physical examination total 3462 groups of effective data samples, a model training sample is set to 2417 groups (69.8%), and a model test sample is set to 1045 groups (30.2%);

s2, standardizing input sample data of each layer according to different input variable dimensions in each layer of the input layer, the hidden layer and the output layer;

s3, setting contents of dependent variables (input layers) as thyroid nodule and thyroid nodule; the factor setting content is such as age, blood pressure and the like; the covariate setting content is quantitative data such as high density lipoprotein cholesterol and the like; establishing a self-learning multi-layer perceptron neural network model;

and S4, drawing ROC curves of a binary logistic regression model and a multi-layer perceptron neural network model respectively by taking the cut-off value calculated by the predictive model formula as a check variable and taking whether the thyroid nodule exists as a state variable.

The test results are:

1) Specific model structure: the input layer is composed of 9 neurons of age, blood pressure and high density lipoprotein cholesterol; the hidden layer is the deviation of each layer, and 6 neurons are all arranged; the output layer was 2 neurons total with and without thyroid nodules. See table 3 below for details.

Table 3 neural network model structure of multilayer perceptron

2) And (3) observing a prediction graph: referring to fig. 5, fig. 5 is an observation prediction diagram of example 3. And cross-comparing sample changes of each partition by using the established multi-layer perceptron neural network model to form an observation prediction graph of thyroid nodules, and defaulting to 0.5 as a correct and error probability demarcation for judging whether the reservoir has thyroid nodules. The ordinate of the predictive graph is the predictive pseudo probability, and the abscissa is the thyroid nodule with and without. According to the observation prediction graph, the reservoir layers are grouped according to the condition that the reservoir layers are suffered from thyroid nodule and are not suffered from thyroid nodule, when any reservoir layer is not suffered from thyroid nodule as a prediction target, the prediction probability of the reservoir layer is obviously higher or lower than the possibility of disease of other reservoir layers, the difference between the reservoir layers is obvious, and the reservoir layer has a good classification and identification effect.

3) Importance variable output: referring to fig. 6, fig. 6 is a comparison chart of importance ranking of input layer parameters according to embodiment 3. The conclusion output by the MLP is mainly influenced by the independent variables of the input layer, and thus the risk factors affecting thyroid nodules are ranked in importance. The earlier single factor analysis results retained 3 risk factors affecting thyroid nodules. From the figure it is seen that blood pressure is an important independent risk factor for thyroid nodule disease.

4) Evaluation effect:

(1) ROC curve

Referring to fig. 7, fig. 7 is a graph of the model ROC of example 3. Area under the curve auc=0.585 of the neural network model of the multi-layer perceptron and area under the curve auc=0.507 of the logistic regression model. The AUC of the multi-layer perceptron neural network model is greater.

(2) Evaluation index

The binary logistic regression model has 95% CI of 0.490-0.523, about dengue index maximum of 0.04324, sensitivity of 12.08% and specificity of 92.248%, and cut-off value of 0.4293, namely, the probability of thyroid nodule is high when the cut-off value is larger than or equal to 0.4293. The multilayer perceptron neural network evaluation model has 95% CI of 0.568-0.601, ROC curve about log index maximum value of 0.1274, sensitivity of 59.76%, specificity of 52.98%, and cut-off value of 0.597, namely when the cut-off value is more than or equal to 0.597, the probability of suffering from thyroid nodule is high. Details are given in table 4 below.

Table 4 comparison of the evaluation ability of the two models

Example 4

The embodiment combines fatty liver in high-incidence diseases of people to construct a crowd health assessment method based on a multi-layer perceptron neural network model, comprising the following steps:

s1, a physical examination total 3462 groups of effective data samples, a model training sample is set to 2417 groups (69.8%), and a model test sample is set to 1045 groups (30.2%);

s2, standardizing input sample data of each layer according to different input variable dimensions in each layer of the input layer, the hidden layer and the output layer;

s3, setting contents of dependent variables (input layers) as fatty liver and non-fatty liver; the factor setting content is such variables as age, sex, BMI, blood pressure and the like; the covariate setting content is quantitative data such as triglyceride, high density lipoprotein cholesterol and the like; and establishing a self-learning multi-layer perceptron neural network model.

And S4, drawing ROC curves respectively drawn into a binary logistic regression model and a multi-layer perceptron neural network model by taking the cut-off value calculated by the predictive model formula as a test variable and taking whether the fatty liver is in the state variable.

The test results are:

1) Specific model structure: the input layer is composed of 16 neurons of age, sex, BMI, blood pressure, triglyceride and high density lipoprotein cholesterol; the hidden layer is the deviation of each layer, and 6 neurons are all arranged; the output layer is composed of 2 neurons with fatty liver and without fatty liver. See table 5 below for details.

Table 5 neural network model structure of multi-layer perceptron

2) And (3) observing a prediction graph: referring to fig. 8, fig. 8 is an observation prediction diagram of example 4. And cross-comparing sample changes of each partition by using the established multi-layer perceptron neural network model to form an observation prediction graph of the fatty liver, and defaulting to 0.5 as a correct and error probability demarcation for judging whether the reservoir is suffering from the fatty liver. The ordinate of the predictive graph is the predictive pseudo-probability, and the abscissa is the presence and absence of fatty liver. According to the result of observing the prediction graph, the reservoir layers are grouped according to the occurrence and nonoccurrence of fatty liver, when any reservoir layer is not affected with fatty liver as a prediction target, the prediction probability of the fatty liver is obviously higher or lower than the possibility of other reservoir layers, the difference between the fatty liver and the reservoir layer is obvious, and the fatty liver classification and identification effect is good.

3) Importance variable output: referring to fig. 9, fig. 9 is a comparison chart of importance ranking of input layer parameters according to embodiment 4. The conclusion output by the MLP is mainly influenced by the independent variables of the input layer, so that importance ranking is carried out on risk factors influencing fatty liver. The earlier single factor analysis results retain 6 risk factors affecting fatty liver. From the figure it is seen that BMI, triglycerides, blood pressure, high density lipoprotein cholesterol are important independent risk factors for fatty liver disease.

4) Evaluation effect:

(1) ROC curve

Referring to fig. 10, fig. 10 is a graph of the model ROC of example 4. Area under the model curve auc=0.722 of the multi-layer perceptron neural network, and area under the model curve auc=0.675 of the logistic regression. The AUC of the multi-layer perceptron neural network model is greater.

(2) Evaluation index

The binary logistic regression model has 95% CI of 0.660-0.691, about dengue index maximum of 0.2579, sensitivity of 76.24 percent and specificity of 49.55 percent, and the cut-off value is 0.58155, namely, when the cut-off value is larger than or equal to 0.58155, the probability of fatty liver is high. The predicted model of the neural network of the multi-layer perceptron has 95% CI of 0.706-0.736, the ROC curve about log index maximum value of 0.3305, the sensitivity of 67.01 percent and the specificity of 66.04 percent, and the cut-off value is 0.387 at the moment, namely, when the cut-off value is more than or equal to 0.387, the possibility of fatty liver is high, and the details are shown in the following table 6.

Table 6 comparison of predictive ability of two models

Example 5

The embodiment provides a health evaluation method based on a multi-layer perceptron neural network model, which comprises the following steps:

s1, constructing a simulated neural network structure model, which comprises the following steps:

s11, respectively dividing the training sample number and the test sample number according to the training test proportion;

s12, performing simulation calculation on relations between diseases and influence factors in different crowds by using training samples;

s13, learning and training among a limited number of samples, and eliminating the correlation and complexity among variables;

s2, training a sample iteration model, and preliminarily obtaining a final model, wherein the step comprises the following steps:

s21, readjusting the health status and influencing factors in the health evaluation process of different people by adopting a decision-making technology, a prediction technology and a simulation technology, and controlling the management mode in real time;

s22, feeding back the human health data information under various conditions in real time;

s23, carrying out iterative verification on the simulated neural network structure model by using a test sample;

s24, after reaching the meaning rule of the satisfactory approximation integral sample, preliminarily obtaining a final model;

s3, optimizing the overall model evaluation effect, including:

s31, optimizing a smoothing factor of a network to improve the accuracy of an evaluation mode in the process of training a sample iteration model;

s32, performing simulation experiment verification comparison, and reducing the number of neurons in the hidden layer;

s33, complexity of the model is reduced, and the overall model evaluation effect is optimized.

In one preferred embodiment, the number of training samples and the number of test samples are obtained from a medical database.

Example 6

The embodiment provides a method for determining a health evaluation standard index, which comprises the following steps:

s1, selecting sensitivity, specificity, cut-off value, prediction accuracy, about dengue index, AUC, ROC curve and 95% credible interval as evaluation standards;

s2, performing visual presentation comparison by using the observation prediction graph and the importance variable ranking graph;

s3, cross-comparing the partitioned samples by using the established multi-layer perceptron neural network model, wherein the abscissa of the partitioned samples is whether the partitioned samples are ill, and the ordinate of the partitioned samples is the prediction pseudo probability;

s4, observing a multi-layer perceptron neural network model in the prediction graph, and defaulting to 0.5 as a probability boundary for judging whether the reservoir has chronic diseases or not.

Example 7

Referring to fig. 11, fig. 11 is a schematic diagram of a health evaluation model of a health evaluation system and method based on a multi-layer sensor neural network model according to the present invention.

The present embodiment provides a health assessment mode, including:

s1, selecting an evaluation index, which comprises the following steps:

s11, medical staff study interviews to know the psychological and social support condition of the corresponding masses;

s12, collecting crowd physical examination data;

s13, knowing indexes of possible chronic disease conditions of the crowd by combining with quantitative blood lipid indexes, blood pressure indexes, electrocardiogram and the like of medical science of hospitals;

s14, analyzing main influencing factors and disease development trend according to the policy file;

s2, data acquisition processing comprises the following steps:

s21, performing intelligent grabbing on specific index data of each azimuth to form a medical database;

s22, carrying out data cleaning, data processing and analysis on a database by combining a multi-layer perceptron neural network algorithm in a data mining technology;

s23, quantifying corresponding indexes;

s3, determining an evaluation connotation, which comprises the following steps:

s31, based on the data processing result and the importance difference of the database, combining the opinions of the specialist in the literature and the Delphi method, and performing reconfirmation;

s32, continuously adjusting the content and the framework of the health assessment, and forming a corresponding large database;

s4, establishing an evaluation mode, which comprises the following steps:

s41, performing overall process management on the steps;

s42, carrying out multi-round verification by combining the problems generated in the clinical practice process;

s43, after repeated correction and iteration, a novel health assessment mode is constructed.

The foregoing is merely a preferred embodiment of the present invention and it should be noted that modifications and additions to the present invention may be made by those skilled in the art without departing from the principles of the present invention and such modifications and additions are to be considered as well as within the scope of the present invention.

Claims (9)

1. The health evaluation system based on the multi-layer perceptron neural network model is characterized by comprising a medical database and a multi-layer perceptron neural network algorithm model, wherein the medical database is used for health evaluation, and the multi-layer perceptron neural network algorithm model is used for operation of a complex database.

2. The system of claim 1, wherein the medical database comprises a sample set and basic information and medical indexes of health data of different people, the sample set comprises general demographics, disease conditions, physiological parameters of human body and other parameters, and the basic information comprises: age, sex, height, weight, BMI, etc.; the medical index includes: physiological parameters, biochemical parameters, blood lipid index, electrocardiogram frequency, etc.

3. The multi-layer perceptron neural network model-based health assessment system of claim 1, characterized in that said multi-layer perceptron neural network algorithm model comprises the steps of: s01, determining the number of hidden layers; s02, determining the number of hidden layer nodes; s03, determining a learning rate factor alpha; s04, selecting a momentum factor beta; s05, selecting an activation function; s06, determining a training test proportion; s07, determining an optimization algorithm.

4. The health evaluation system based on the multi-layer sensor neural network model according to claim 3, wherein the main conditions for the calculation of the number of explicitly hidden layer nodes in step S02 are: the sum of the total node number must be less than N-1, N being the number of training samples; the connection weight of the network structure model is smaller than the number of training samples, and the multiple between the connection weight and the training samples is 2-10 times.

5. The health evaluation system based on a multi-layer perceptron neural network model of claim 3, wherein in the iterative process, the new learning rate is the original learning rate multiplied by a constant Φ, expressed as: α (t+1) =α (t) ×Φ, and when the total error function VE < 0, the value range of the constant Φ is: phi is more than 0.001 and less than 1; when the total error function VE > 0, the value range of the constant phi is as follows: phi is more than 1 and less than 50.

6. The health evaluation system based on the multi-layer perceptron neural network model of claim 3, wherein in the step S04, the previous weight change is used to smooth the learning path in the process of adjusting and modifying the weights, and the expression is: w (t+1) =w (t) +βvw (t), wherein W (t) is a main weight in the t-th iteration, βvw (t) is a direction of adjustment change of the memory t weight, and a main range of values of the constant β is 0-t.

7. The health evaluation system based on the multi-layer perceptron neural network model of claim 3, wherein the Sigmoid function expression used in step S05 is:

the parameter A is used for adjusting the overall change trend and range of data, and promoting the change rate of the original gentle region to be continuously accelerated, and the value range is as follows: a is more than 0.01 and less than 100; the parameter B is generally 1, so that the function sensitivity is higher; the parameter C is used for eliminating an insensitive interval in the data, continuously increasing a change area of the data, and the value range is as follows: c is more than 1 and less than 100.

8. A method of health assessment using a multi-layer perceptron neural network model of any of claims 1-7, the method comprising:

s1, constructing a simulated neural network structure model, which comprises the following steps:

s11, respectively dividing the training sample number and the test sample number according to the training test proportion;

s12, performing simulation calculation on relations between diseases and influence factors in different crowds by using training samples;

s13, learning and training among a limited number of samples, and eliminating the correlation and complexity among variables;

s2, training a sample iteration model, and preliminarily obtaining a final model, wherein the step comprises the following steps:

s21, readjusting the health status and influencing factors in the health evaluation process of different people by adopting a decision-making technology, a prediction technology and a simulation technology, and controlling the management mode in real time;

s22, feeding back the human health data information under various conditions in real time;

s23, carrying out iterative verification on the simulated neural network structure model by using a test sample;

s24, after reaching the meaning rule of the satisfactory approximation integral sample, preliminarily obtaining a final model;

s3, optimizing the overall model evaluation effect, including:

s31, optimizing a smoothing factor of a network to improve the accuracy of an evaluation mode in the process of training a sample iteration model;

s32, performing simulation experiment verification comparison, and reducing the number of neurons in the hidden layer;

s33, complexity of the model is reduced, and the overall model evaluation effect is optimized.

9. The method of claim 8, wherein the training sample number and the test sample number are obtained from a medical database.

Images