Heat stroke disease risk assessment method based on big data and physical sign monitoring

Abstract

The present application relates to the technical field of neural network model, and particularly relates to a heat stroke disease risk assessment method based on big data and sign monitoring, which comprises the following steps: dividing samples in a sign data set into several subsets, and assigning a disease risk label to each subset; using all the subsets and the disease risk labels to perform a first training of a neural network, wherein a dropout parameter of the neural network in the first training is set as a; obtaining samples with a disease risk label mismatch according to an output result of the neural network after the first training; changing the disease risk label of the samples with the disease risk label mismatch; and performing a second training of the neural network, wherein a dropout parameter of the neural network in the second training is set as a1, and a1 is less than a; and using the neural network after the second training to perform a heat stroke disease risk assessment. The training method improves the accuracy of the neural network model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of neural network models, and in particular to a heat stroke disease risk assessment method based on big data and physical sign monitoring. BACKGROUND

[0002] Heat stroke is a serious heat stress disease, usually occurring in extremely high temperature environments, when the human body cannot effectively regulate body temperature. This condition is the most severe form of heatstroke, and if not treated promptly, it can lead to multiple organ failure and even more serious consequences, so it is very important to prevent heat stroke or assess the risk of heat stroke.

[0003] The common method for assessing the risk of heat stroke is to detect physical indicators such as body temperature, heart rate, blood pressure, blood oxygen, and pulse, and to determine whether they are outside the set range of values. When the set data range is exceeded, it is determined that the risk of disease is high, and a warning is issued. However, considering that different people have different physical conditions and underlying diseases, the occurrence of symptoms also varies greatly, and this method has a large error in disease risk assessment. In addition, neural networks can also be used to assess the risk of disease, but on the one hand, there is a lack of labeled or accurately labeled training samples, making it difficult for neural networks to be adequately trained; on the other hand, the physical condition and underlying disease of different people are also different, and the occurrence of symptoms also varies greatly, resulting in low accuracy of neural network training methods. SUMMARY

[0004] To solve the above problems, the present application provides a heat stroke disease risk assessment method based on big data and physical sign monitoring.

[0005] The heat stroke disease risk assessment method based on big data and physical sign monitoring of the present application adopts the following technical solutions:

[0006] One embodiment of the present application provides a heat stroke disease risk assessment method based on big data and physical sign monitoring, which includes the following steps:

[0007] Obtain a physical sign dataset, and the physical sign data of each sample in the physical sign dataset includes body temperature, heart rate, blood pressure, and blood oxygen.

[0008] The sample in which the heat stroke disease occurs is recorded as a sick sample, the samples in the sign data set are divided into several subsets according to the numerical range of the sign data of the sick sample, each subset is assigned a sick risk label according to the distribution of the sign data of the samples in each subset outside the numerical range, the first training of the neural network is performed by using all the subsets and their sick risk labels, the dropout parameter of the neural network is set to a preset value a during the first training, the samples with mismatched sick risk labels are obtained according to the output result of the neural network after the first training, the sick risk labels of the samples with mismatched sick risk labels are changed, and then the second training of the neural network is performed, the dropout parameter of the neural network is set to a1 during the second training, and a1 is less than a, and the neural network after the second training is used for heat stroke sick risk assessment.

[0009] Preferably, the step of dividing the samples in the sign data set into several subsets according to the numerical range of the sign data of the sick sample comprises the following specific steps:

[0010] The maximum value max and the minimum value min of each sign data of all sick samples are obtained, and [min, max] is taken as the numerical range of each sign data of the sick sample;

[0011] The numerical range of each sign data is equally divided into several subintervals, and [0, min) and (max, +∞) are also taken as two subintervals, respectively;

[0012] Each subinterval is numbered in the order from left to right;

[0013] The number of the subinterval to which each sign data of each sample in the sign data set belongs is obtained, the number corresponding to each sign data of each sample is recorded as the distribution feature of each sample, and the samples with the same distribution feature are taken as a subset.

[0014] Preferably, the step of assigning a sick risk label to each subset according to the distribution of the sign data of the samples in each subset outside the numerical range comprises the following specific steps:

[0015] The difference between each sign data of each sample in each subset and the left endpoint of the numerical range of the sign data is recorded as the deviation index of each sign data of each sample, and the sick risk label of each subset is obtained according to the mean value of the deviation indexes of all sign data of all samples in each subset.

[0016] Preferably, the step of obtaining the samples with mismatched sick risk labels according to the output result of the neural network after the first training comprises the following specific steps:

[0017] All the sign data of each sample in the sign data set is input into the neural network after the first training, and the difference between the neural network output result and the disease risk label of the subset to which each sample belongs is calculated, which is recorded as the disease risk error of each sample; for the disease risk errors of all samples in the sign data set, the sample corresponding to the outlier disease risk error is recorded as the sample with a mismatched disease risk label.

[0018] Preferably, the method for changing the disease risk label of the sample with a mismatched disease risk label comprises the following specific steps:

[0019] The subset S1 to which the sample with a mismatched disease risk label belongs is obtained, and the disease risk label of the subset S1 is recorded as L1;

[0020] When the disease risk error of the sample with a mismatched disease risk label is less than th1, the disease risk label that is less than L1 and has the smallest difference with L1 is taken as the disease risk label of the sample with a mismatched disease risk label among the disease risk labels of all subsets; th1 is a preset value less than or equal to 0;

[0021] When the disease risk error of the sample with a mismatched disease risk label is greater than th2, the disease risk label that is greater than L1 and has the smallest difference with L1 is taken as the disease risk label of the sample with a mismatched disease risk label among the disease risk labels of all subsets; th2 is a preset value greater than or equal to 0.

[0022] Preferably, the calculation method of a1 is as follows:

[0023] The disease risk label mismatch coefficient w is calculated according to the outlier disease risk error;

[0024] ;

[0025] ;

[0026] wherein represents an exponential function with a natural constant as the base, and .

[0027] Preferably, the disease risk label mismatch coefficient w is equal to the mean value of the absolute values of all the outlier disease risk errors.

[0028] Preferably, the method for obtaining the outlier disease risk error comprises:

[0029] The disease risk errors of all samples in the sign data set are input into the LOF algorithm to obtain the outlier disease risk error.

[0030] Preferably, the loss function used by the neural network in the first training and the second training is a mean square error loss function.

[0031] Preferably, the input of the neural network is all the sign data of each sample.

[0032] The beneficial effects of the technical solutions of the present application are: according to the numerical range of the sign data of the diseased sample, the samples in the sign data set are divided into several subsets, and according to the distribution of the sign data of the samples in each subset within the numerical range, a disease risk label is assigned to each subset. Each subset can reflect the approximate distribution of the sign data when the heat stroke symptom occurs, and each subset can be analyzed independently without being disturbed and affected by too many other samples. Therefore, a disease risk label can be assigned to each subset to generally describe the possible disease risk of each subset, avoiding the situation that the neural network cannot be fully and accurately trained due to the lack of labels of the samples in the sign data set.

[0033] The first training of the neural network is performed using all the subsets and their disease risk labels, and the dropout parameter of the neural network is set to a preset value a during the first training. The disease risk labels of the samples that do not match the disease risk labels are changed based on the output results of the neural network after the first training, and the second training of the neural network is performed. The dropout parameter of the neural network is set to a1 during the second training, and a1 is less than a. Since the disease risk labels used in the first training of the neural network can only roughly reflect the occurrence of the heat stroke symptom under different samples, that is, the first training of the neural network should learn the general and shallow heat stroke disease features of the sign data set, and it does not need to learn too much about the heat stroke disease features of the samples with incorrect labels. Based on the neural network after the first training, the disease risk labels are changed and corrected, and a1 is less than a to ensure that the neural network can further learn the deep disease features of all samples in the sign data, which includes the symptom features of different physical conditions and underlying diseases of different people, thereby improving the accuracy of the neural network training. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0035] Figure 1 The step flow chart of the heat stroke disease risk assessment method based on big data and sign monitoring provided by an embodiment of the present application. DETAILED DESCRIPTION

[0036] In order to further clarify the technical means and effects taken by the present application to achieve the predetermined inventive purpose, the following describes in detail the specific implementation, structure, features and effects of the heat stroke disease risk assessment method based on big data and sign monitoring according to the present application, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0038] The following describes in detail the specific scheme of the heat stroke disease risk assessment method based on big data and sign monitoring provided by the present application in combination with the accompanying drawings. Embodiment one

[0039] Please refer to Figure 1 which shows the step flowchart of the heat stroke disease risk assessment method based on big data and sign monitoring provided by one embodiment of the present application, which includes the following steps:

[0040] Step S1, obtaining a sign data set.

[0041] The heat stroke susceptible population wears a sign detection device, such as a bracelet that can be used for sign monitoring, which can monitor sign data such as body temperature, heart rate, blood pressure, blood oxygen, number of pulses per minute, blood sugar, and other embodiments can also monitor environmental data such as environmental temperature and humidity, and also incorporate environmental data into feature data; in this embodiment, the sign detection device collects sign data every second as a time point. And the sign detection device can transmit all sign data at each time point to the cloud database and store it.

[0042] The heat stroke susceptible population includes athletes, firefighters, construction workers engaged in outdoor labor, the elderly and children, etc. in high temperature weather. When a person has heat stroke symptoms and is treated, the person's sign data is labeled as abnormal. Generally, sign detection devices can only detect sign data and cannot directly determine whether symptoms occur, so when symptoms occur, other personnel, such as medical personnel, first aid personnel or sick personnel, send an abnormality through the sign detection device to perform abnormal labeling, such as pressing the button for sending an abnormal signal in the sign detection device.

[0043] It should be noted that not all personnel with heat stroke symptoms are marked as abnormal, because in actual production and life, when heat stroke symptoms occur, the focus is on eliminating symptoms or emergency treatment, so there is no time or obligation to mark as abnormal.

[0044] In this embodiment, all feature data of each person at each time in the database is taken as a sample, and in other embodiments, all feature data of each person within a certain period of time can be taken as a sample, for example, all sign data of each person within one minute before each time is taken as a sample.

[0045] When the sign data of the personnel with heat stroke symptoms is marked as abnormal, not all feature data at each time needs to be marked as abnormal, because the timing of marking as abnormal may not be the timing of heat stroke symptoms, or even when marked as abnormal, the personnel with heat stroke symptoms has recovered and the sign data has returned to normal, or the symptoms are more serious.

[0046] This embodiment only marks part of the sign data of the personnel with heat stroke symptoms as abnormal; for example: assuming that personnel A has heat stroke symptoms, the average value of all feature data of personnel A at each time is obtained, which is recorded as the average sign at each time; in other embodiments, the average value of all feature data of personnel A within a period of time before each time is obtained, which is also recorded as the average sign at each time. Then all feature data of the N time points with the largest average sign is marked as abnormal, this embodiment takes N=60 as an example for description, other embodiments can be set to other values, and this embodiment will not be described.

[0047] In other embodiments, for all time points within the day when personnel A has heat stroke symptoms, the average sign at each time is divided into two parts using the Otsu threshold segmentation algorithm, and the sign data at the time point with the largest average value of the average sign is marked as abnormal.

[0048] In other embodiments, considering that all sign data of each personnel at each time needs to be stored, which leads to a large amount of useless data (such as normal sign data of personnel at rest) stored in the database within a certain period of time, for example, one week or one month; in order to reduce the large amount of storage of useless data, only the sign data collected at all time points within the last five days can be stored, and the data collected at other time points is deleted, but the sign data marked as abnormal is not deleted.

[0049] In summary, the database contains a number of samples, which constitute a sign data set, part of which are labeled as abnormal, and are referred to as sick samples. At the same time, it is considered that not all sign data of personnel with heat stroke symptoms are labeled as abnormal, which leads to the fact that samples other than sick samples may also have heat stroke symptoms, that is, some samples in the sign data set lack labels.

[0050] Step S2, dividing the samples in the sign data set into a plurality of subsets according to the numerical range of the sign data of the sick samples, and assigning a sick risk label to each subset according to the distribution of the sign data of the samples in each subset within the numerical range.

[0051] Some samples in the above-obtained sign data set have symptoms but are not labeled as abnormal, and the above labeling is only a rough approximation (for example, the above labeling of all feature data of the N time points with the largest average sign as abnormal), which leads to the fact that part of the data in the sign data set lacks labels, and the labeled samples may also have false labeling.

[0052] The sign data set with missing labels or false labels is difficult to fully train an accurate neural network for evaluating or predicting heat stroke.

[0053] For some data sets lacking labels, although a small sample learning method can be used to train a neural network, the samples in the sign data set in the present embodiment come from different personnel, and the physical condition or underlying diseases of different personnel are different, leading to different timing or conditions of heat stroke symptoms, which makes it impossible to accurately train a neural network using a small sample learning method.

[0054] Therefore, the present embodiment needs to reassign labels to the sign data set with missing labels and false labels.

[0055] Step S201, dividing the samples in the sign data set into a plurality of subsets according to the numerical range of the sign data of the sick samples.

[0056] It is considered that the physical condition or underlying diseases of different personnel lead to a large amount of noise in the samples in the sign data set. Therefore, the samples in the sign data set are divided into a plurality of subsets, and each subset is analyzed independently without interference and influence from other samples.

[0057] It is also considered that the sign data of the existing sick samples can generally reflect the sign characteristics when heat stroke symptoms appear, so the samples in the sign data set are divided into a plurality of subsets according to the numerical range of the sign data of the sick samples.

[0058] As a preferred example, the method for obtaining the subsets is as follows:

[0059] Obtain the maximum value max and the minimum value min of each sign data of all sick samples, and take [min, max] as the numerical range of each sign data of the sick samples.

[0060] The numerical range of each sign data is equally divided into several sub-intervals, and each sub-interval is numbered in order from left to right; in this embodiment, 10 sub-intervals are divided, and in other embodiments, other numbers can be divided, which are not limited in this embodiment.

[0061] The numerical range of each sign data is equally divided into several sub-intervals, and [0, min) and (max, +∞) are also respectively taken as two sub-intervals.

[0062] Obtain the number of the sub-interval to which each sign data of each sample in the sign data set belongs, and the number corresponding to all sign data of each sample is recorded as the distribution feature of each sample. Samples with the same distribution feature are taken as a subset.

[0063] The above method of obtaining each subset can ensure that the samples in each subset have approximate sign data distribution, and different subsets have different feature distribution. Different subsets can be used to roughly distinguish the sign features of heat stroke in different sick conditions. Subsequent analysis of each subset can not be disturbed and affected by other samples.

[0064] In other embodiments, considering that some people may have heat stroke symptoms due to physical fitness, but their sign data may not fall within the above numerical range, which leads to insufficient division of the above subsets. At this time, the numerical range can be appropriately expanded, for example, the numerical range [min, max] is expanded to [min-L, max+L], where L can be set to (max-min) x 10%. Then divide the sub-intervals and obtain several subsets based on the expanded numerical range.

[0065] Step S202, according to the distribution of the sign data of the samples in each subset outside the numerical range, assign a sick risk label to each subset.

[0066] Each subset obtained above is divided based on the distribution range of the sign data of the sick samples, so each subset can reflect the approximate distribution of the sign data when heat stroke symptoms occur, and each subset can be analyzed without too much interference and influence from other samples. Therefore, a sick risk label can be roughly assigned to each subset to describe the possible sick risk of each subset as a whole, avoiding the situation that the neural network cannot be fully and accurately trained due to the lack of labels of the sign data set samples.

[0067] As an example, the method of assigning a disease risk label to each subset comprises:

[0068] The difference between each sign data of each sample in each subset and the left end point (i.e. min) of the numerical range of the sign data is recorded as the deviation index of each sign data of each sample. The disease risk label of each subset is obtained according to the mean of the deviation indexes of all sign data of all samples in each subset. Note that the disease risk labels of all subsets need to be linearly normalized, and the disease risk labels used subsequently are linearly normalized disease risk labels.

[0069] The above process considers that the greater the deviation index of a sample feature, the more likely it is abnormal. In order to generally describe the disease risk from the whole, the mean of the deviation indexes of all sign data is used to obtain the disease risk label of each subset.

[0070] One possible way is to directly use the deviation index of each sign data of each sample as the disease risk label of each sample, so that each sample can be directly assigned a label (disease risk label) to describe the disease risk of each sample, solving the problem of lack of labels in the sign data set. However, this approach cannot solve the interference problem caused by different physical conditions or underlying diseases of different personnel.

[0071] Therefore, the embodiment does not directly assign a label to each sample, but assigns a label (disease risk label) to each subset, that is, to generally describe the disease risk from the whole, to some extent to avoid the interference caused by different physical conditions or underlying diseases of different personnel.

[0072] Step S3, using all subsets and their disease risk labels to perform the first training of the neural network, and setting the dropout of the neural network to a preset value a during the first training.

[0073] Each subset corresponds to a disease risk label, and in the embodiment, the label of each sample is set to the disease risk label of the subset to which the sample belongs. The first training of the neural network is performed using all samples in the sign data set and their disease risk labels.

[0074] The input of the neural network is all feature data of each sample, the neuron parameters of the neural network are randomly initialized, the parameter update method uses the stochastic gradient descent method, and the loss function used during training is the mean square error loss function; the specific training process of the neural network is known, and the embodiment will not be described in detail.

[0075] The neural network used in this embodiment is a fully connected neural network with 4 hidden layers. In other embodiments, other structures of fully connected neural networks or other types of neural networks, such as LSTM neural networks, can be used. The structure and type of the neural network are not specifically limited in this embodiment.

[0076] Since the disease risk label used in the first training of the neural network can only roughly reflect the occurrence of heatstroke symptoms in different samples, that is, the neural network should learn the general and shallow heatstroke disease features of the general public in the first training, and it is not necessary to learn the heatstroke disease features of the samples with annotation errors. Therefore, the dropout of the neural network in the first training should be set to be large. In this embodiment, the dropout parameter is set to a, and a=0.5 is taken as an example for description.

[0077] Step S4: obtaining samples with mismatched disease risk labels according to the output results of the neural network after the first training.

[0078] Since the neural network after the first training can extract general and shallow heatstroke disease features in the sign data set, samples with annotation errors and mismatched disease risk labels can be obtained according to the output results of the neural network after the first training.

[0079] As an example, the method for obtaining samples with mismatched disease risk labels includes:

[0080] Each sample in the sign data set is input into the neural network after the first training, and the difference between the output result of the neural network and the disease risk label of each sample is calculated, which is recorded as the disease risk error of each sample. For the disease risk errors of all samples in the sign data set, the sample corresponding to the outlier disease risk error is recorded as the sample with mismatched disease risk label.

[0081] As a preferred example, the method for obtaining the outlier disease risk error includes:

[0082] The disease risk errors of all samples in the sign data set are input into the LOF algorithm to obtain the outlier disease risk error.

[0083] As an optional example, the method for obtaining the outlier disease risk error includes:

[0084] The mean shift clustering algorithm is used to cluster the disease risk errors of all samples in the sign data set to obtain a plurality of categories, and the disease risk errors outside all categories are taken as the outlier disease risk error.

[0085] Step S5, changing the disease risk label of the sample with the disease risk label mismatch, and performing the second training of the neural network, the dropout setting of the neural network in the second training is a1, a1 is less than a.

[0086] In order to accurately extract the disease features of the sample with the disease risk label mismatch, the label of the sample needs to be changed so as to enable the neural network to learn more heatstroke features in the sign data.

[0087] As an example, the method for changing the disease risk label of the sample with the disease risk label mismatch includes the following steps.

[0088] Obtaining a subset S1 to which the sample with the disease risk label mismatch belongs, and the disease risk label of the subset S1 is L1.

[0089] When the disease risk error of the sample with the disease risk label mismatch is less than th1, in all disease risk labels of the subsets, the disease risk label less than L1 and having the least difference with L1 is taken as the disease risk label of the sample with the disease risk label mismatch.

[0090] Wherein th1 is a value less than or equal to 0, when the disease risk error of the sample with the disease risk label mismatch is less than th1, it indicates that the disease risk label of the sample is assigned larger, and therefore the disease risk label less than L1 and having the least difference with L1 is taken as the disease risk label of the sample with the disease risk label mismatch. The present embodiment is described by taking th1 =-0.2 as an example, other embodiments can be set to other values, for example, set to 0, and the present embodiment is not specifically limited.

[0091] When the disease risk error of the sample with the disease risk label mismatch is greater than th2, in all disease risk labels of the subsets, the disease risk label greater than L1 and having the least difference with L1 is taken as the disease risk label of the sample with the disease risk label mismatch.

[0092] Wherein th2 is a value greater than or equal to 0, when the disease risk error of the sample with the disease risk label mismatch is greater than th2, it indicates that the disease risk label of the sample is assigned smaller, and therefore the disease risk label greater than L1 and having the least difference with L1 is taken as the disease risk label of the sample with the disease risk label mismatch. The present embodiment is described by taking th2 =0.2 as an example, other embodiments can be set to other values, for example, set to 0, and the present embodiment is not specifically limited. If there is no disease risk label less than L1 and having the least difference with L1, or there is no disease risk label greater than L1 and having the least difference with L1, the disease risk label of the sample with the disease risk label mismatch is not changed at this time.

[0093] After the disease risk labels of the samples with mismatched disease risk labels are changed, the neural network is trained for the second time using the disease risk labels of all samples in the sign data.

[0094] In the second training, the neuron parameters of the neural network are initialized as the parameters of the neural network after the first training, the parameter updating method is the stochastic gradient descent method, and the loss function used in the training is the mean square error loss function. The specific training process of the neural network is known, and will not be described in detail in this embodiment.

[0095] After the risk labels of the samples with mismatched disease risk labels are changed, the disease risk labels of all samples in the sign data are more accurate than before, so in the second training, the neural network needs to further learn the deep disease characteristics of all samples in the sign data, which includes the manifestation of disease characteristics of different people with different physical conditions and underlying diseases. Therefore, the dropout of the neural network in the second training should be set smaller, and in this embodiment, it is set to a1. This embodiment is described taking a1=0.25 as an example.

[0096] Step S6, using the neural network trained for the second time to evaluate the heatstroke disease risk.

[0097] All sign data of each person collected in real time at the current time is input into the neural network trained for the second time. The result output by the neural network is recorded as a heatstroke disease risk evaluation index. When the heatstroke disease risk evaluation index is greater than a threshold th3, a warning is given, for example, a warning signal is sent to the sign monitoring device, and the sign monitoring device gives a voice prompt after receiving the warning signal, reminding the person of the risk of heatstroke disease. This embodiment is described taking th3=0.7 as an example, and other values can be set according to specific conditions in other embodiments, which will not be limited in this embodiment. Embodiment two

[0098] As a preferred example, the method for obtaining the dropout parameter a1 further comprises:

[0099] The disease risk label mismatch coefficient w is calculated according to the outlying disease risk error.

[0100] ;

[0101] ;

[0102] wherein represents the exponential function with the natural constant as the base, and .

[0103] Wherein the disease risk label mismatch coefficient w describes whether the disease risk label mismatch is serious, when w is larger, it means that the disease risk label mismatch is serious, there are a large number of samples and their disease risk labels do not match, or a large number of samples are mislabeled, or there is a part of the sample whose actual assigned label and the real label are quite different; At this time, during the second training, the dropout parameter a1 is smaller than a, but it cannot be too small compared with a, to avoid the problem that the second training of the neural network deeply learns the wrong features caused by the serious disease risk label mismatch.

[0104] When w is smaller, it means that the disease risk label mismatch is not serious, most of the sample labels are correct or approximately correct; At this time, during the second training, the dropout parameter a1 can be much smaller than a, to ensure that the second training of the neural network deeply learns the correct features.

[0105] Therefore, the above preferred example further automatically adjusts the dropout parameter a1 on the basis of changing the risk label of the sample with disease risk label mismatch, so as to ensure the accuracy and reliability of the second training of the neural network, that is, the disease characteristics contained in the physical data are deeply extracted and learned, and the deep learning of the wrong features is avoided in the learning process. Embodiment three

[0106] After the second training of the neural network is completed, the samples with disease risk label mismatch are obtained again according to the output results of the neural network after the second training, the disease risk labels of the samples with disease risk label mismatch are changed, and the third training of the neural network is performed, the dropout parameter of the neural network during the second training is set to a2, and a2 is smaller than a1. The calculation method of a2 is: replacing a with a1 and replacing a1 with a2 in the calculation method of embodiment two, and the specific calculation method is the same as that in embodiment two.

[0107] After the third training of the neural network is completed, the samples with disease risk label mismatch are obtained again according to the output results of the neural network after the third training, the disease risk labels of the samples with disease risk label mismatch are changed, and the fourth training of the neural network is performed, and so on, until the proportion of the samples with disease risk label mismatch is less than 10%, or the M times of training are stopped, and this embodiment takes M=10 as an example for description. Embodiment four

[0108] In the case where the neural network is trained multiple times, for example, 2 to 10 times, if there is no diseased risk label less than L1 and least different from L1, or no diseased risk label greater than L1 and least different from L1, the diseased risk label of the sample with a mismatched diseased risk label is reduced or increased by 10%, until the diseased risk label is less than 0 or greater than 1.

[0109] The above description is merely preferred embodiments of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application should be included in the protection scope of the present application.

Claims

1. A heat stroke disease risk assessment method based on big data and vital sign monitoring, characterized in that, The method comprises the following steps: Obtaining a sign data set, the sign data of each sample in the sign data set comprising: body temperature, heart rate, blood pressure, blood oxygen; Wherein the sample with heat stroke disease is recorded as a sick sample, the samples in the sign data set are divided into several subsets according to the numerical range of the sign data of the sick sample, each subset is assigned a sick risk label according to the distribution of the sign data of the samples in each subset within the numerical range, the first training of the neural network is performed by using all the subsets and their sick risk labels, the dropout parameter of the neural network is set to a preset value a during the first training, the sick risk label mismatched sample is obtained according to the output result of the neural network after the first training, the sick risk label of the sick risk label mismatched sample is changed, and then the second training of the neural network is performed, the dropout parameter of the neural network is set to a1 during the second training, and a1 is less than a, the second trained neural network is used for heat stroke sick risk assessment; The specific steps of dividing the samples in the sign data set into several subsets according to the numerical range of the sign data of the sick sample comprise the following steps: Obtaining the maximum value max and the minimum value min of each sign data of all sick samples, and taking [min, max] as the numerical range of each sign data of the sick sample; The numerical range of each sign data is equally divided into several subintervals, and [0, min) and (max, +∞) are also taken as two subintervals, respectively; Each subinterval is numbered in the order from left to right; Obtaining the number of the subinterval to which each sign data of each sample in the sign data set belongs, the number corresponding to each sample is recorded as the distribution characteristics of each sample, and the samples with the same distribution characteristics are taken as a subset; The specific steps of assigning a sick risk label to each subset according to the distribution of the sign data of the samples in each subset within the numerical range comprise the following steps: The difference between each sign data of each sample in each subset and the left endpoint of the numerical range of the sign data is recorded as the deviation index of each sign data of each sample, and the mean value of the deviation indexes of all sign data of all samples in each subset is obtained to obtain the sick risk label of each subset. 2.The heat stroke disease risk assessment method based on big data and vital sign monitoring according to claim 1, wherein, The specific steps of obtaining the sick risk label mismatched sample according to the output result of the neural network after the first training comprise the following steps: All sign data of each sample in the sign data set is input into the neural network after the first training, the difference between the output result of the neural network and the sick risk label of the subset to which each sample belongs is calculated, which is recorded as the sick risk error of each sample, and the sample corresponding to the sick risk error of the outlier is recorded as the sick risk label mismatched sample. 3.The heat stroke disease risk assessment method based on big data and vital sign monitoring according to claim 2, wherein, The specific steps of changing the sick risk label of the sick risk label mismatched sample comprise the following steps: Obtaining the subset S1 to which the sick risk label mismatched sample belongs, and the sick risk label of the subset S1 is recorded as L1; When the disease risk error of the sample with the mismatched disease risk label is less than th1, in all disease risk labels of the subsets, the disease risk label less than L1 and having the least difference with L1 is taken as the disease risk label of the sample with the mismatched disease risk label; th1 is a preset value less than or equal to 0; When the disease risk error of the sample with the mismatched disease risk label is greater than th2, in all disease risk labels of the subsets, the disease risk label greater than L1 and having the least difference with L1 is taken as the disease risk label of the sample with the mismatched disease risk label; th2 is a preset value greater than or equal to 0. 4.The heat stroke disease risk assessment method based on big data and vital sign monitoring according to claim 2, wherein, The calculation method of a1 is: The disease risk label mismatching coefficient w is calculated according to the outlier disease risk error; ; ; wherein denotes the exponential function with base the natural constant, and . 5.The heat stroke disease risk assessment method based on big data and vital sign monitoring according to claim 4, wherein, The disease risk label mismatching coefficient w is equal to the mean value of the absolute values of all outlier disease risk errors. 6.The heat stroke disease risk assessment method based on big data and vital sign monitoring according to claim 2, wherein, The method for obtaining the outlier disease risk error comprises: The disease risk errors of all samples in the sign data set are input into the LOF algorithm to obtain the outlier disease risk error. 7.The heat stroke disease risk assessment method based on big data and vital sign monitoring according to claim 1, wherein, The loss function adopted by the neural network in the first training and the second training is a mean square error loss function. 8.The heat stroke disease risk assessment method based on big data and vital sign monitoring according to claim 1, wherein, The input of the neural network is all sign data of each sample.

Images