Generative adversarial auto-encoder-based healthy population standard distribution model construction method

Abstract

The invention provides a healthy crowd standard distribution model construction method based on a generative adversarial auto-encoder. The method comprises the steps of obtaining a first data set; the first data set is a data set containing electronic health record data of healthy people; preprocessing the first data set; dividing the preprocessed first data set into a training set and a test set; building a network model of the generative adversarial auto-encoder, and carrying out weight initialization operation on the model; constructing a reconstruction loss function, a crowd representativeness constraint loss function and an adversarial loss function; training and optimizing the model by using the preprocessed training set, and storing an optimal model; and inputting the preprocessed test set into the optimal model to predict the standard distribution of the healthy people, and obtaining a prediction result of the standard distribution of the healthy people. According to the method, the distribution of ideal reference individuals can be obtained, and meanwhile, the generated healthy population standard distribution prediction result is ensured to have certain rationality and authenticity.

Description

Technical Field

[0001] The present invention relates to the technical field of standard distribution of healthy people, and in particular to a method for constructing a standard distribution model of healthy people based on a generative adversarial autoencoder. Background Art

[0002] Reference intervals play a vital role in the interpretation of laboratory data and can help clinicians form a more comprehensive perspective in diagnosis and treatment management. On the one hand, establishing a suitable comparable cohort is essential for accurately interpreting patient data. Patient data, including comparisons with reference intervals, will be used to support, question or exclude clinical diagnoses, and may also discover unrecognized diseases, which is an important contribution of the laboratory to improving overall health. However, the currently published reference data often have shortcomings. For example, the cohorts described may lack the same standards as reference interval studies, and the descriptions of clinical severity and duration are subjective. Studies with small sample sizes may not accurately reflect the diversity of the population. On the other hand, even with reference intervals determined based on suitable comparable cohorts, diagnosis is still a serious problem for patients with common diseases in the "gray area", because over-interpretation of laboratory or clinical information may lead to anxiety, increased medical costs and other hazards.

[0003] In order to determine an accurate and applicable reference interval, it is first necessary to select a representative healthy population as a reference population. This population should be as similar as possible to the target patient population in terms of demographic characteristics such as age and gender. There are different methods for selecting reference individuals. For large-scale multi-level studies, there are suggestions for selecting reference individuals based on non-clinical criteria, but this requires a lot of funding to support data collection. In some smaller studies, referral medical institutions or hospitals are relied on to provide reliable diagnoses, and diagnostic information can be obtained in about 85% of cases, so that suitable reference individuals can be screened out in this way. It is very important to select an appropriate reference population, and its test results must be clinically relevant. When selecting a healthy population as a reference, factors such as age, gender and region should be considered. Reference intervals are usually established based on data from healthy populations, but data from people of different regions, genders and ages may vary significantly, which means that the universal applicability of reference intervals may be limited, especially when dealing with marginal cases or special circumstances. In addition, the testing conditions and methods of different laboratories may vary, which may lead to differences in the results of different laboratories even within the same reference interval. With the development of medical science and technology, new detection technologies and methods are constantly emerging, and existing reference intervals may need to be updated to reflect the latest clinical practices and technological advances.

[0004] Among them, how to select a representative healthy population as a reference is a complex process, which has brought obstacles to the study of reference intervals and the determination of suitable comparable cohorts. In other words, defining a reference population is a complex process that requires comprehensive consideration of sample selection, data processing, application of statistical methods, and regional and population specificity. Through these steps, a scientific and practical reference interval can be established to provide an important decision support tool for clinical diagnosis. Usually, it is possible to comprehensively judge whether a patient belongs to the healthy population in the medical reference interval through preliminary screening, clinical evaluation, and expert consultation. Specifically, preliminary screening includes questionnaires and physical examinations. Clinical evaluation is to comprehensively assess the patient's health status in combination with the patient's clinical symptoms and other relevant examination results (such as electrocardiograms, liver function tests, etc.). Expert consultation is for patients with doubts or special circumstances, and professional doctors should be sought to ensure the accuracy of the diagnosis. Through the above steps, it can be more comprehensively judged whether the patient belongs to the healthy population in the medical reference interval. In addition, when selecting a reference population to establish a reference interval, some inclusion and exclusion criteria can be formulated to take into account some factors that may affect the accuracy of the reference interval.

[0005] At present, there are two main methods for establishing reference intervals: indirect method and direct method. The main difference between the indirect method and the direct method is that the indirect method does not directly collect samples from selected healthy populations to determine the reference interval, but uses existing data sets that may contain a wider population. The advantages of this method include low cost, fast process, and no inconvenience or discomfort to patients. However, a major limitation of the indirect method is that it may be affected by disease subgroups, which may lead to deviations in the reference interval. When using the indirect method to establish a reference interval, it is first necessary to collect a large amount of historical data, which usually comes from the Laboratory Information System (LIS) or other databases. Then, the data is processed by statistical analysis methods such as skewness-kurtosis test, BOX-COX transformation, interquartile range method and Hoffmann method to eliminate outliers and calculate the reference interval. In addition, the influence of factors such as gender and age needs to be considered, as these factors may affect the levels of certain biomarkers. When selecting a reference population, it is necessary to ensure that the selected population is representative and that known disease states are excluded as much as possible. For example, if the reference interval is used to assess male fertility, the most appropriate reference population would be men who have successfully conceived a pregnancy with their spouse within the past year. In addition, the indirect method can also be stratified according to different biological variables, which helps to define the reference interval more accurately. In summary, the indirect method provides an efficient and low-cost method for establishing reference intervals, especially when it is not possible to collect enough samples directly from the target population. However, in order to ensure the accuracy and applicability of the reference interval, the data must be carefully selected and processed, and all factors that may affect the results must be taken into account.

[0006] However, the selection of healthy people in the above existing methods for establishing reference intervals obviously has the following defects:

[0007] (1) The representativeness of the reference population is difficult to ensure. The best practices for selecting healthy people as reference populations nationwide should include reasonable geographical distribution, strict screening criteria, scientific data collection methods, methodological processes that follow international standards, and multi-channel data collection strategies. The reference sample group should cover major regions of my country to ensure that the data is widely representative. However, the cost of determining normal / abnormal labels for patients is high.

[0008] (2) The standard distribution of healthy population is unknown, and there is a lack of such research reports. Therefore, there is no prior knowledge available for determining the reference interval of healthy population.

[0009] (3) The criteria for defining healthy reference individuals are highly subjective. When screening reference individuals, a comprehensive assessment is required through questionnaires, physical examinations, laboratory tests, and imaging examinations. Specific requirements include being aware of one's health and having no obvious chronic diseases or other conditions that affect health. The judgment is highly subjective. Selecting a representative healthy population as a reference population is an important challenge currently facing the industry.

[0010] With the advent of the big data era, deep learning technology has significant advantages in processing ultra-large-scale medical data. By using these technologies, large amounts of medical data can be analyzed and processed more comprehensively. Therefore, we need to develop a method for constructing a standard distribution model of healthy people based on deep learning, so as to quickly construct a standard reference population, and then determine the medical reference interval range for disease diagnosis more economically and conveniently. Summary of the invention

[0011] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0012] To this end, the purpose of the present invention is to propose a method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder.

[0013] In order to achieve the above-mentioned purpose, the technical solution of the present invention provides a method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder, the method comprising: step S1: obtaining a first data set; wherein the first data set is a data set containing electronic health record data of a healthy population; step S2: preprocessing the first data set; step S3: dividing the preprocessed first data set into a training set and a test set according to a preset ratio; step S4: building a network model of a generative adversarial autoencoder and performing a weight initialization operation on the network model; wherein the network model comprises a standard population prior distribution generator, an encoder, a decoder and a discriminator; the standard population prior distribution generator is used to generate a prior distribution of a standard population based on a selected prior distribution function; wherein the prior distribution of the standard population is used to simulate the standard distribution of the population composition of the target population applied to the network model; the encoder is used to map the input data X to a hidden variable Z; wherein the input data X is a preprocessed training set or a test set; the hidden variable Z is a latent space representation generated by the encoder; the decoder is used to map the hidden variable Z back to a feature matrix X' of the same dimension as the input data; The discriminator takes the latent variable Z and the latent space representation Z' sampled from the prior distribution of the standard population as input, is used to distinguish the latent variable Z from the latent space representation Z' sampled from the prior distribution of the standard population, and simultaneously outputs a probability value from 0 to 1, wherein the probability value represents the probability that the latent variable Z is considered by the discriminator to be from the latent space representation Z'; Step S5: constructing a reconstruction loss function, a population representativeness constraint loss function, and an adversarial loss function; the reconstruction loss function is used to measure the difference between the input data X and the feature matrix X'; the population representativeness constraint The loss function is used to measure the similarity between the feature matrix X' and the target standard population distribution; the adversarial loss function is used to measure the difference between the latent variable Z and the potential space representation Z' sampled from the prior distribution of the standard population; step S6: based on the reconstruction loss function, the population representativeness constraint loss function and the adversarial loss function, and using the preprocessed training set to train and optimize the network model, save the optimal model; step S7: input the preprocessed test set into the optimal model to predict the standard distribution of the healthy population, and obtain the prediction result of the standard distribution of the healthy population.

[0014] Preferably, the method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder also includes: step S8: obtaining a second data set; wherein the second data set is a data set obtained in real time and contains electronic health record data of certain patients; step S9: preprocessing the second data set; step S10: inputting the preprocessed second data set Y into the saved optimal model to obtain a feature matrix Y' of the same dimension as the second data set output by the decoder; step S11: calculating the mean square error value of the preprocessed second data set Y and the feature matrix Y'; step S12: determining whether the mean square error value is greater than or equal to a preset threshold; when the determination result is yes, the second data set is determined to be abnormal data.

[0015] Preferably, the step S6 specifically includes: step S6.1: inputting the preprocessed training set into the encoder of the network model after weight initialization to generate a latent variable Z; step S6.2: inputting the latent variable Z into the decoder of the network model after weight initialization to generate a feature matrix X' of the same dimension as the preprocessed training set; step S6.3: performing gradient descent on the error value corresponding to the reconstruction loss function to synchronously update the weights of the encoder and decoder at this time until the error value corresponding to the reconstruction loss function converges; step S6.4: performing gradient descent on the error value corresponding to the adversarial loss function to force the latent variable Z generated by the encoder to be close to the potential space representation sampled from the prior distribution of the standard population composition. Show Z', and synchronously update the weights of the discriminator and encoder at this time until the error value corresponding to the adversarial loss function converges; and step S6.5: after training each batch, extract statistical features from the generated samples; the statistical features are then compared with the standard demographic features to calculate the value of the population representativeness constraint loss function; through this comparison, the output value of the population representativeness constraint loss function will be updated; the training process will continue until the output value of the population representativeness loss function reaches the preset target value; finally, the generative adversarial autoencoder network model is optimized in this way to ensure that the generative adversarial autoencoder network model can fully reflect the representative characteristics of the population when generating samples; wherein the generated samples are the feature matrix X'.

[0016] Preferably, the selected prior distribution function is a Gaussian distribution with a mean value of 0 and a standard deviation of 1.

[0017] Preferably, the electronic health record data includes: clinical measurement data; wherein, the clinical measurement data includes one of the following or a combination thereof: demographic data, physical examination data, laboratory examination data, imaging examination data; the demographic data includes one of the following or a combination thereof: gender, age, height, weight, body mass index, waist circumference; the physical examination data includes one of the following or a combination thereof: diastolic blood pressure, systolic blood pressure; the laboratory examination data includes one of the following or a combination thereof: blood examination data, urine examination data; the imaging examination data includes one of the following or a combination thereof: examination data in an X-ray examination report, examination data in a CT examination report, examination data in an MRI examination report, examination data in an ultrasound examination report.

[0018] Preferably, the expression of the reconstruction loss function is:

[0019]

[0020] In formula (1), x represents the input sample of the encoder, that is, the input data X; x′ represents the output sample of the decoder, that is, the feature matrix X′; N represents the number of samples in the data set; i represents the i-th input sample of the encoder; x i represents the data value of the i-th input sample of the encoder; x′ i represents the data value of the i-th output sample of the decoder;

[0021] The expression of the adversarial loss function is:

[0022]

[0023] In formula (2), z represents a random variable sampled from the latent space representation sampled from the prior distribution of the standard population; p z represents the distribution of the random variable z; D(z) represents the probability value output by the discriminator; Represents the distribution p z The expected value of z is obtained by sampling; z′ is the output obtained by inputting real data into the generator; q z is the distribution related to the generator; D(z′) is the discriminator’s judgment result on z′;

[0024] The population representativeness constraint loss function includes measurement indicators at the feature level and the overall level;

[0025] Among them, the representative score at the feature level is defined as It is used to represent the similarity of generated data on population feature k. The overall representative score is defined as gl, which is used to measure the comprehensive similarity of generated data on all features.

[0026] The expression of the representative constraint loss function is as follows:

[0027]

[0028] In formula (3), EP trial represents the set of generated samples that satisfy all constraints; EP pool represents a set of samples that meet any single feature condition; if the generated data cannot meet the representativeness requirements of all features, a weighted feature-level constraint is introduced, and the expression is as follows:

[0029]

[0030] In formula (4), K is the total number of features; w k is the importance weight of feature k; is the representative score at the feature level.

[0031] The total loss function combines reconstruction loss, adversarial loss and representative constraint loss. The total loss function is as follows:

[0032]

[0033] In formula (5), α and β are loss weight coefficients.

[0034] Preferably, the step S2 specifically includes: step S2.1: performing data cleaning on the first data set; step S2.2: performing compressed sensing on the first data set after data cleaning; step S2.3: normalizing the first data set after compressed sensing; step S2.4: performing data enhancement on the first data set after normalization; and / or the step S9 specifically includes: step S9.1: performing data cleaning on the second data set; step S9.2: performing compressed sensing on the second data set after data cleaning; step S9.3: normalizing the second data set after compressed sensing; step S9.4: performing data enhancement on the second data set after normalization.

[0035] Beneficial effects of the present invention:

[0036] The method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder provided by the present invention can obtain the distribution of an ideal reference individual: an individual in a healthy state can be used as an ideal reference individual. At the same time, the method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder provided by the present invention can ensure that the prediction results of the generated standard distribution of a healthy population have certain rationality and authenticity.

[0037] The representative constraint loss is integrated into the total loss function of the generative adversarial autoencoder. By adjusting these two weights during the training process, the best compromise can be found between data reproduction (fidelity), generative adversarial (distribution regularization), and representativeness (matching the overall statistical characteristics of the reference population).

[0038] However, in actual clinical practice, it is often necessary to find other individuals with similar characteristics to the current patient to construct a comparable cohort for comparison and analysis. The present invention avoids the gap between subjective description and objective standards in determining reference intervals and constructing comparable cohorts, as well as the lack of unified standards and specifications, and provides an important tool for establishing a unified and accurate reference interval database and clinical application guidelines.

[0039] In addition, the method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder provided by the present invention can not only screen the standard distribution of a healthy population, but also achieve the purpose of screening a diseased population.

[0040] Additional aspects and advantages of the invention will become apparent from the following description, or may be learned by practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic flow chart of a method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder according to an embodiment of the present invention is shown;

[0042] Figure 2 A schematic diagram of the architecture of a network model for generating an adversarial autoencoder according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0043] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, Figure 1 and Figure 2 As shown, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementations. It should be noted that, in the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0044] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0045] Figure 1 A schematic flowchart of a method for constructing a standard distribution model for a healthy population based on a generative adversarial autoencoder according to an embodiment of the present invention is shown. Figure 2FIG. 1 is a schematic diagram showing the overall architecture of a network model of a generative adversarial autoencoder constructed according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder includes:

[0046] Step S1: obtaining a first data set; wherein the first data set is a data set containing electronic health record data of healthy people;

[0047] Step S2: preprocessing the first data set;

[0048] Step S3: Divide the preprocessed first data set into a training set and a test set according to a preset ratio;

[0049] Step S4: constructing a network model for generating adversarial autoencoders and performing weight initialization operations on the network model; wherein the network model includes a standard population prior distribution generator, an encoder, a decoder, and a discriminator;

[0050] A standard population prior distribution generator, used to generate a standard population prior distribution based on a selected prior distribution function; wherein the standard population prior distribution is used to simulate the standard distribution of the population composition of the target population when the network model is applied;

[0051] An encoder is used to map input data X to a latent variable Z; wherein the input data X is a preprocessed training set or test set; and the latent variable Z is a latent space representation generated by the encoder;

[0052] A decoder, used to map the latent variable Z back to a feature matrix X' of the same dimension as the input data;

[0053] a discriminator, taking the latent variable Z and the latent space representation Z' sampled from the prior distribution of the standard population as input, for distinguishing the latent variable Z from the latent space representation Z' sampled from the prior distribution of the standard population, and simultaneously outputting a probability value from 0 to 1, wherein the probability value represents the probability that the latent variable Z is considered by the discriminator to be from the latent space representation Z';

[0054] Step S5: constructing a reconstruction loss function, a crowd representativeness constraint loss function, and an adversarial loss function;

[0055] A reconstruction loss function, used to measure the difference between the input data X and the feature matrix X';

[0056] The crowd representativeness constraint loss function is used to measure the similarity between the feature matrix X' and the target standard crowd distribution;

[0057] An adversarial loss function that measures the difference between the latent variable Z and the latent space representation Z' sampled from the prior distribution of the standard population;

[0058] Step S6: Based on the reconstruction loss function, the crowd representativeness constraint loss function and the adversarial loss function, the network model is trained and optimized using the preprocessed training set, and the optimal model is saved;

[0059] Step S7: Input the preprocessed test set into the optimal model to predict the standard distribution of the healthy population, and obtain the prediction result of the standard distribution of the healthy population.

[0060] In this embodiment, the method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder provided by the present invention can obtain the distribution of an ideal reference individual: an individual in a healthy state can be used as an ideal reference individual. At the same time, the method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder provided by the present invention can ensure that the prediction results of the generated standard distribution of a healthy population have certain rationality and authenticity.

[0061] However, in actual clinical practice, it is often necessary to find other individuals with similar characteristics to the current patient to construct a comparable cohort for comparison and analysis. The present invention avoids the gap between subjective description and objective standards in determining reference intervals and constructing comparable cohorts, as well as the lack of unified standards and specifications, and provides an important tool for establishing a unified and accurate reference interval database and clinical application guidelines.

[0062] Furthermore, the method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder provided by the present invention chooses to build a network model of a generative adversarial autoencoder. On the one hand, it fits the real-life scenario where reference individuals lack objective labels and electronic medical record data is sparse; on the other hand, a generative adversarial neural network is an unsupervised learning method that can be trained on a large amount of unlabeled health data to learn the distribution pattern of health data, thereby obtaining a model that can identify individuals with abnormal diseases. Since the definition and standard distribution of a healthy population is a complex and multidimensional concept, it involves multiple aspects such as physical, psychological, and social adaptability. Understanding the definition and characteristics of a healthy population, as well as taking corresponding management recommendations, is of great significance for maintaining and improving personal health levels. However, the definition of a healthy population lacks objective evaluation criteria, and the diagnostic reference interval of a medical disease usually needs to rely on the distribution of a healthy population to be defined. However, the definition of a healthy population is very subjective and lacks a unified quantitative standard. Since there are actually no objective labels that can be used for supervised machine learning modeling. Therefore, we choose to build a network model of a generative adversarial autoencoder:

[0063] First of all, the network model of the generative adversarial autoencoder we built mainly works in an unsupervised manner and does not require a large amount of labeled data, which is especially useful in the medical field where data is scarce.

[0064] Secondly, the network model of the generative adversarial autoencoder can generate highly realistic high-quality data mainly due to its unique adversarial training mechanism. The encoder constantly tries to generate samples close to the real data distribution, while the discriminator strives to distinguish between real data and generated data. Through this continuous adversarial, the encoder gradually improves the quality of generated samples, making it difficult to distinguish the generated samples from real data. The discriminator distinguishes between diseased and healthy people as much as possible, and at the same time uses the population distribution rules specific to the medical field (i.e., the latent space representation Z' sampled from the prior distribution of the standard population) as a constraint to ensure that the generated data has a certain rationality and authenticity. Through continuous adversarial training, the encoder gradually learns to simulate the real data distribution of healthy people.

[0065] Furthermore, in this embodiment, we describe the specific application of the standard population prior distribution generator in practice. Since the standardization of population structure is an important statistical method, it involves converting population data from different regions, different age groups, and different genders into data of a standard population for comparison. China's standard population composition mainly includes age, gender, urban and rural distribution, ethnic composition, and education level. This method helps to eliminate the impact of population structure differences on statistical data, so as to more accurately evaluate and analyze population characteristics and their impact on group health indicators. The acquisition of China's standard population structure mainly relies on methods such as the national census, population sampling surveys, and population change sampling surveys. These methods ensure the comprehensiveness and rationality of the generated data.

[0066] In a specific embodiment, we use the standard population distribution of the Chinese population as a prior distribution to constrain the population data distribution structure fitted by our generative adversarial autoencoder network model, avoiding the underfitting and overfitting of the model caused by the difference between the population structure and the standard population distribution. The prior knowledge of the standard population distribution data is used to guide the data generation process to generate reasonable population distribution data, so that our generative adversarial autoencoder network model learns to generate samples that meet the standard population distribution of the Chinese population.

[0067] In one embodiment of the present invention, the method for constructing a standard distribution model for a healthy population based on a generative adversarial autoencoder also includes: step S8: obtaining a second data set; wherein the second data set is a data set obtained in real time and contains electronic health record data of certain patients; step S9: preprocessing the second data set; step S10: inputting the preprocessed second data set Y into the saved optimal model to obtain a feature matrix Y' of the same dimension as the second data set output by the decoder; step S11: calculating the mean square error value of the preprocessed second data set Y and the feature matrix Y'; step S12: determining whether the mean square error value is greater than or equal to a preset threshold; when the determination result is yes, the second data set is determined to be abnormal data.

[0068] In one embodiment of the present invention, the step S6 specifically includes: step S6.1: inputting the preprocessed training set into the encoder of the network model after weight initialization to generate a hidden variable Z; step S6.2: inputting the hidden variable Z into the decoder of the network model after weight initialization to generate a feature matrix X' of the same dimension as the preprocessed training set; step S6.3: performing gradient descent on the error value corresponding to the reconstruction loss function to synchronously update the weights of the encoder and decoder at this time until the error value corresponding to the reconstruction loss function converges; step S6.4: performing gradient descent on the error value corresponding to the adversarial loss function to force the hidden variable Z generated by the encoder to be close to the latent variable sampled from the prior distribution of the standard population composition. Represent Z' in space, and synchronously update the weights of the discriminator and encoder at this time until the error value corresponding to the adversarial loss function converges; and step S6.5: after training each batch, extract statistical features from the generated samples; the statistical features are then compared with the standard demographic features to calculate the value of the population representativeness constraint loss function; through this comparison, the output value of the population representativeness constraint loss function will be updated; the training process will continue until the output value of the population representativeness loss function reaches the preset target value; finally, the generative adversarial autoencoder network model is optimized in this way to ensure that the generative adversarial autoencoder network model can fully reflect the representative characteristics of the population when generating samples; wherein the generated samples are the feature matrix X'.

[0069] In one embodiment of the present invention, the selected prior distribution function is a Gaussian distribution with a mean value of 0 and a standard deviation of 1.

[0070] In one embodiment of the present invention, the electronic health record data includes: clinical measurement data; wherein the clinical measurement data includes one of the following or a combination thereof: demographic data, physical examination data, laboratory examination data, and imaging examination data; the demographic data includes one of the following or a combination thereof: gender, age, height, weight, body mass index, and waist circumference; the physical examination data includes one of the following or a combination thereof: diastolic blood pressure and systolic blood pressure; the laboratory examination data includes one of the following or a combination thereof: blood examination data and urine examination data; the imaging examination data includes one of the following or a combination thereof: examination data in an X-ray examination report, examination data in a CT examination report, examination data in an MRI examination report, and examination data in an ultrasound examination report.

[0071] In one embodiment of the present invention, the expression of the reconstruction loss function is:

[0072]

[0073] In formula (1), x represents the input sample of the encoder, that is, the input data X; x′ represents the output sample of the decoder, that is, the feature matrix X′; N represents the number of samples in the data set; i represents the i-th input sample of the encoder; x i represents the data value of the i-th input sample of the encoder; x′ i represents the data value of the i-th output sample of the decoder;

[0074] The expression of the adversarial loss function is:

[0075]

[0076] In formula (2), z represents a random variable sampled from the latent space representation sampled from the prior distribution of the standard population; p z represents the distribution of the random variable z; D(z) represents the probability value output by the discriminator; Represents the distribution p z The expected value of z is obtained by sampling; z′ is the output obtained by inputting real data into the generator; q z is the distribution related to the generator; D(z′) is the discriminator’s judgment result on z′;

[0077] The population representativeness constraint loss function includes measurement indicators at the feature level and the overall level;

[0078] Among them, the representative score at the feature level is defined as It is used to represent the similarity of generated data on population feature k. The overall representative score is defined as gl, which is used to measure the comprehensive similarity of generated data on all features.

[0079] The expression of the representative constraint loss function is as follows:

[0080]

[0081] In formula (3), EP trial represents the set of generated samples that satisfy all constraints; EP pool represents a set of samples that meet any single feature condition; if the generated data cannot meet the representativeness requirements of all features, a weighted feature-level constraint is introduced, and the expression is as follows:

[0082]

[0083] In formula (4), K is the total number of features; w k is the importance weight of feature k; is the representative score at the feature level.

[0084] The total loss function combines reconstruction loss, adversarial loss and representative constraint loss. The total loss function is as follows:

[0085]

[0086] In formula (5), α and β are loss weight coefficients.

[0087] In one embodiment of the present invention, the step S2 specifically includes: step S2.1: performing data cleaning on the first data set; step S2.2: performing compressed sensing on the first data set after data cleaning; step S2.3: normalizing the first data set after compressed sensing; step S2.4: performing data enhancement on the first data set after normalization; and / or the step S9 specifically includes: step S9.1: performing data cleaning on the second data set; step S9.2: performing compressed sensing on the second data set after data cleaning; step S9.3: normalizing the second data set after compressed sensing; step S9.4: performing data enhancement on the second data set after normalization.

[0088] The following is a specific example to illustrate the technical solution of the present invention. The method for constructing a standard distribution model of healthy people based on a generative adversarial autoencoder of the present invention is specifically implemented as follows:

[0089] (1) Obtain a first data set; wherein the first data set is a data set containing electronic health record data of healthy people; in fact, we use the electronic health record data of 6315 healthy people.

[0090] The electronic health record data includes: clinical measurement data;

[0091] Among them, the clinical measurement data includes one of the following or a combination thereof: demographic data, physical examination data, laboratory examination data, and imaging examination data; the demographic data includes one of the following or a combination thereof: gender, age, height, weight, body mass index, and waist circumference; the physical examination data includes one of the following or a combination thereof: diastolic blood pressure, systolic blood pressure; the laboratory examination data includes one of the following or a combination thereof: blood examination data, urine examination data; the imaging examination data includes one of the following or a combination thereof: examination data in an X-ray examination report, examination data in a CT examination report, examination data in an MRI examination report, and examination data in an ultrasound examination report.

[0092] (2) Preprocessing the first data set; Preprocessing the first data set, specifically including: cleaning the first data set; performing compressed sensing on the first data set after data cleaning; normalizing the first data set after compressed sensing; and performing data enhancement on the first data set after normalization. We solve the problem of data sparsity through compressed sensing operations, reconstruct the complete first data set through a small number of measurements, and improve the efficiency and performance of the network model of the generated adversarial autoencoder.

[0093] Specifically, the application of compressed sensing in the data preprocessing stage: before inputting into the encoder, the high-dimensional clinical data is compressed using compressed sensing technology. The data is projected into a low-dimensional space through a compressed sensing matrix. This process can significantly reduce the problem of data unavailability caused by the missing dimension of the input data X, and improve the training efficiency of the network model of the subsequent generative adversarial autoencoder.

[0094] (3) Divide the preprocessed first data set into a training set and a test set according to an 8:2 ratio.

[0095] (4) Building a network model for generating adversarial autoencoders and performing weight initialization on the network model; Figure 2 As shown, the network model includes a standard population prior distribution generator, an encoder, a decoder, and a discriminator;

[0096] A standard population prior distribution generator is used to generate a prior distribution of a standard population based on a selected prior distribution function; wherein the prior distribution of the standard population is used to simulate the standard distribution of the population composition of the target population when the network model is applied; an encoder is used to map input data X to a latent variable Z; wherein the input data X is a preprocessed training set or test set; the latent variable Z is a latent space representation generated by the encoder; a decoder is used to map the latent variable Z back to a feature matrix X' (i.e. Figure 2The discriminator takes the latent variable Z and the latent space representation Z' sampled from the prior distribution of the standard population as input, is used to distinguish the latent variable Z from the latent space representation Z' sampled from the prior distribution of the standard population, and simultaneously outputs a probability value from 0 to 1, wherein the probability value represents the probability that the latent variable Z is considered by the discriminator to be from the latent space representation Z'.

[0097] Furthermore, we introduce the structure of the encoder (6-layer neural network), as follows:

[0098] First convolutional layer: kernel size 3×3, stride 1, output channels 32, LeakyReLU activation function. First max pooling layer: pooling window size 2×2, stride 2, feature map dimensionality reduction. Second convolutional layer: kernel size 3×3, stride 1, output channels 64, Leaky ReLU activation function. Second max pooling layer: pooling window size 2×2, stride 2, further dimensionality reduction. First fully connected layer: flatten features and map to hidden layer representation. Second fully connected layer: map hidden layer to latent space representation.

[0099] Further, we introduce the structure of the decoder, which is as follows: The first fully connected layer: maps the latent space back to the hidden representation. The second fully connected layer: decodes back to the flattened feature map. The deconvolution layer: gradually restores the feature map to the same size as the input. The output layer: generates the reconstructed data using the Sigmoid activation function (i.e. Figure 2 The generated data X' in .

[0100] Since the adversarial network consists of a generator and a discriminator: the generator (the generative adversarial autoencoder we built does not contain a generator alone, the encoder essentially plays the role of a generator). Further, we introduce the discriminator: using a multi-layer fully connected network, inputting real low-dimensional data or generated data, and outputting true or false probabilities.

[0101] Decoder output: Get the reconstructed low-dimensional data.

[0102] In this step, we can also use the reconstruction data generated by the decoder (i.e. Figure 2 The generated data X' in the image is combined with the reconstructed low-dimensional data) and the compressed sensing algorithm to restore the high-dimensional data. Specifically, the compressed sensing algorithm: uses optimization algorithms such as L1 regularization to reconstruct high-dimensional data from low-dimensional representation.

[0103] The selected prior distribution function is a Gaussian distribution with a mean of 0 and a standard deviation of 1. Specifically, we simulate the standard distribution of healthy people based on existing electronic health record (EHR) data.

[0104] (5) Constructing a reconstruction loss function, a crowd representativeness constraint loss function, and an adversarial loss function; wherein the reconstruction loss function is used to measure the difference between the input data X and the feature matrix X'; the crowd representativeness constraint loss function is used to measure the similarity between the feature matrix X' and the target standard population distribution; the adversarial loss function is used to measure the difference between the latent variable Z and the latent space representation Z' sampled from the prior distribution of the standard population. We use the adversarial loss function, the crowd representativeness constraint loss function, and the reconstruction loss function to form a total loss to optimize the encoder and the discriminator.

[0105] Among them, the expression of the reconstruction loss function is:

[0106]

[0107] In formula (1), x represents the input sample of the encoder, that is, the input data X; x′ represents the output sample of the decoder, that is, the feature matrix X′; N represents the number of samples in the data set; i represents the i-th input sample of the encoder; x i represents the data value of the i-th input sample of the encoder; x′ i represents the data value of the i-th output sample of the decoder;

[0108] The expression of the adversarial loss function is:

[0109]

[0110] In formula (2), z represents a random variable sampled from the latent space representation sampled from the prior distribution of the standard population; p z represents the distribution of the random variable z; D(z) represents the probability value output by the discriminator; Represents the distribution p z The expected value of z is obtained by sampling; z′ is the output obtained by inputting real data into the generator; q z is the distribution related to the generator; D(z′) is the discriminator’s judgment result on z′;

[0111] In order to further improve the consistency of the generated data in the distribution of population characteristics, a population representativeness constraint loss function is added.

[0112] The population representativeness constraint loss function includes measurement indicators at the feature level and the overall level;

[0113] Among them, the representative score at the feature level is defined as It is used to represent the similarity of generated data on population feature k. The overall representative score is defined as gl, which is used to measure the comprehensive similarity of generated data on all features.

[0114] The expression of the representative constraint loss function is as follows:

[0115]

[0116] In formula (3), EP trial represents the set of generated samples that satisfy all constraints; EP pool represents a set of samples that meet any single feature condition; if the generated data cannot meet the representativeness requirements of all features, a weighted feature-level constraint is introduced, and the expression is as follows:

[0117]

[0118] In formula (4), K is the total number of features; w k is the importance weight of feature k; is the representative score at the feature level.

[0119] The total loss function combines reconstruction loss, adversarial loss and representative constraint loss. The total loss function is as follows:

[0120]

[0121] In formula (5), α and β are loss weight coefficients.

[0122] By optimizing the above total loss function, high-fidelity reconstruction of the input data, regularization of the latent space distribution, and representative constraints on the population distribution of the generated data can be achieved, thereby generating high-quality data that meets the standard population distribution. (6) Based on the reconstruction loss function, the population representativeness constraint loss function, and the adversarial loss function, the network model of the generative adversarial autoencoder is trained and optimized using the preprocessed training set, and the optimal model is saved;

[0123] The method is based on the reconstruction loss function, the crowd representativeness constraint loss function and the adversarial loss function, and uses the preprocessed training set to train and optimize the network model of the generative adversarial autoencoder, and save the optimal model, specifically including:

[0124] Input the preprocessed training set into the encoder of the network model after weight initialization to generate a hidden variable Z; input the hidden variable Z into the decoder of the network model after weight initialization to generate a feature matrix X' with the same dimension as the preprocessed training set; perform gradient descent on the error value corresponding to the reconstruction loss function to synchronously update the weights of the encoder and decoder at this time until the error value corresponding to the reconstruction loss function converges; perform gradient descent on the error value corresponding to the adversarial loss function to force the hidden variable Z generated by the encoder to be close to the potential space representation Z' sampled from the prior distribution of the standard population composition, and synchronously update the discriminator and encoder at this time until the error value corresponding to the adversarial loss function converges; and after training each batch, extracting statistical features from the generated samples; then comparing the statistical features with the standard demographic features to calculate the value of the population representativeness constraint loss function; through this comparison, the output value of the population representativeness constraint loss function will be updated; the training process will continue until the output value of the population representativeness loss function reaches a preset target value; finally, the generative adversarial autoencoder network model is optimized in this way to ensure that the generative adversarial autoencoder network model can fully reflect the representative characteristics of the population when generating samples; wherein the generated samples are the feature matrix X'.

[0125] (7) The preprocessed test set is input into the optimal model to predict the standard distribution of the healthy population, and the prediction results of the standard distribution of the healthy population are obtained.

[0126] (8) Acquiring a second data set; wherein the second data set is a data set containing electronic health record data of certain patients acquired in real time; the electronic health record data includes: clinical measurement data;

[0127] Among them, the clinical measurement data includes one of the following or a combination thereof: demographic data, physical examination data, laboratory examination data, and imaging examination data; the demographic data includes one of the following or a combination thereof: gender, age, height, weight, body mass index, and waist circumference; the physical examination data includes one of the following or a combination thereof: diastolic blood pressure, systolic blood pressure; the laboratory examination data includes one of the following or a combination thereof: blood examination data, urine examination data; the imaging examination data includes one of the following or a combination thereof: examination data in an X-ray examination report, examination data in a CT examination report, examination data in an MRI examination report, and examination data in an ultrasound examination report.

[0128] (9) Preprocessing the second data set; Preprocessing the second data set, specifically including: cleaning the second data set; performing compressive sensing on the cleaned second data set; normalizing the compressed sensing second data set; and enhancing the normalized second data set. We solve the problem of data sparsity through compressed sensing operations, reconstruct the complete second data set through a small number of measurements, and improve the efficiency and performance of the network model of the generated adversarial autoencoder.

[0129] Specifically, the application of compressed sensing in the data preprocessing stage: before inputting into the encoder, the high-dimensional clinical data is compressed using compressed sensing technology. The data is projected into a low-dimensional space through a compressed sensing matrix. This process can significantly reduce the problem of data unavailability caused by the missing dimension of the input data X, and improve the training efficiency of the network model of the subsequent generative adversarial autoencoder.

[0130] (10) Inputting the preprocessed second data set Y into the saved optimal model to obtain a feature matrix Y' output by the decoder with the same dimension as the second data set;

[0131] (11) Calculate the mean square error of the preprocessed second data set Y and the feature matrix Y';

[0132] (12) Determine whether the mean square error value is greater than or equal to a preset threshold; when the judgment result is yes, determine that the second data set is abnormal data, thereby achieving the purpose of detecting the presence of abnormal data in the electronic health record data of certain patients.

[0133] (13) We used electronic health record data from 6,315 healthy people to simulate the performance of the above-mentioned generative adversarial autoencoder network model. Specifically, we used precision, recall, and F1 score to evaluate the performance of the network model in detecting normal samples and abnormal samples.

[0134] Among them, precision, recall, and F1 score are commonly used classification evaluation indicators. Their definitions and formulas are as follows:

[0135] 1. Precision: Precision measures the proportion of samples that are actually positive among all samples predicted by the model to be positive. That is, how accurate the model's prediction is. The formula for precision is as follows:

[0136]

[0137] Among them, TP (True Positive) is the number of samples correctly predicted by the model as positive. FP (False Positive) is the number of samples incorrectly predicted by the model as positive but actually negative.

[0138] 2. Recall: The recall rate measures the proportion of samples that the model can correctly identify among all samples that are actually positive. That is, how many positive samples the model can find. The formula for the recall rate is as follows:

[0139]

[0140] Among them, TP (True Positive): the number of samples correctly predicted by the model as positive. FN (False Negative): the number of samples incorrectly predicted by the model as negative but actually positive.

[0141] 3. F1 Score: F1 score is the harmonic mean of precision and recall, which combines the performance of precision and recall. It is particularly suitable for classification problems when the positive and negative samples are unbalanced. The formula of F1 score is as follows:

[0142]

[0143] Among them, the F1 score provides a comprehensive indicator that can be used when a trade-off between precision and recall is needed. Its value is between 0 and 1, and the larger the value, the better the overall performance of the model.

[0144] We use precision, recall, and F1 scores to evaluate the performance of the network model of the generative adversarial autoencoder we built. Experiments show that the network model of the generative adversarial autoencoder we built is very effective in constructing the standard distribution of healthy people, and can obtain ideal reference individuals while ensuring that the prediction results of the generated standard distribution of healthy people have a certain degree of rationality and authenticity. It avoids the gap between subjective description and objective standards in determining reference intervals and building comparable cohorts, as well as the lack of unified standards and specifications, and provides an important tool for establishing a unified and accurate reference interval database and clinical application guidelines.

[0145] In summary, the selection of a clinically relevant reference population is crucial to determining the accurate reference interval of laboratory tests. The current improvement in laboratory precision and accuracy has not directly translated into improved clinical understanding, partly due to the failure to fully utilize existing medical data. In addition, there are issues with the complexity of defining reference intervals and the difficulty of selecting an appropriate reference population. Current methods are all for a single indicator, and the definition of screening and exclusion criteria under the guidance of expert experience is highly subjective. Existing methods are unable to conduct research and tools on methods for constructing standard distributions of reference populations across indicators.

[0146] It has been proved that the method for constructing a standard distribution model of a healthy population based on a generative adversarial autoencoder provided by the present invention has unsurpassed technical advantages in related technologies.

[0147] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for constructing a standard distribution model of healthy people based on generative adversarial autoencoders, characterized in that: include: Step S1: Acquire a first data set; wherein the first data set is a data set containing electronic health record data of healthy people; Step S2: preprocessing the first data set; Step S3: Divide the preprocessed first data set into a training set and a test set according to a preset ratio; Step S4: constructing a network model for generating adversarial autoencoders, and performing a weight initialization operation on the network model; wherein the network model includes a standard population prior distribution generator, an encoder, a decoder, and a discriminator; The standard population prior distribution generator is used to generate a prior distribution of a standard population based on a selected prior distribution function; wherein the prior distribution of the standard population is used to simulate the standard distribution of the population composition of the target population when the network model is applied; The encoder is used to map input data X to latent variables Z; wherein the input data X is a preprocessed training set or test set; the latent variable Z is a latent space representation generated by the encoder; The decoder is used to map the latent variable Z back to a feature matrix X' of the same dimension as the input data; The discriminator, taking the latent variable Z and the latent space representation Z' sampled from the prior distribution of the standard population as input, is used to distinguish the latent variable Z from the latent space representation Z' sampled from the prior distribution of the standard population, and simultaneously outputs a probability value from 0 to 1, wherein the probability value represents the probability that the latent variable Z is considered by the discriminator to be from the latent space representation Z'; Step S5: constructing a reconstruction loss function, a crowd representativeness constraint loss function, and an adversarial loss function; The reconstruction loss function is used to measure the difference between the input data X and the feature matrix X'; The population representativeness constraint loss function is used to measure the similarity between the feature matrix X' and the target standard population distribution; The adversarial loss function is used to measure the difference between the latent variable Z and the latent space representation Z' sampled from the prior distribution of the standard population; Step S6: Based on the reconstruction loss function, the crowd representativeness constraint loss function and the adversarial loss function, the network model is trained and optimized using the preprocessed training set, and the optimal model is saved; Step S7: input the preprocessed test set into the optimal model to predict the standard distribution of the healthy population, and obtain the prediction result of the standard distribution of the healthy population.

2. The method for constructing a standard distribution model of healthy people based on generative adversarial autoencoders according to claim 1, characterized in that: Also includes: Step S8: Acquire a second data set; wherein the second data set is a data set obtained in real time and contains electronic health record data of certain patients; Step S9: preprocessing the second data set; Step S10: inputting the preprocessed second data set Y into the saved optimal model to obtain a feature matrix Y' output by the decoder with the same dimension as the second data set; Step S11: Calculating the mean square error value of the preprocessed second data set Y and the feature matrix Y'; Step S12: determining whether the mean square error value is greater than or equal to a preset threshold; when the determination result is yes, determining that the second data set is abnormal data.

3. The method for constructing a standard distribution model of healthy people based on generative adversarial autoencoders according to claim 1, characterized in that: The step S6 specifically includes: Step S6.1: Input the preprocessed training set into the encoder of the network model after weight initialization to generate latent variables Z; Step S6.2: inputting the latent variable Z into the decoder of the network model after weight initialization to generate a feature matrix X' of the same dimension as the preprocessed training set; Step S6.3: performing gradient descent on the error value corresponding to the reconstruction loss function to synchronously update the weights of the encoder and decoder at this time until the error value corresponding to the reconstruction loss function converges; Step S6.4: performing gradient descent on the error value corresponding to the adversarial loss function to force the latent variable Z generated by the encoder to be close to the latent space representation Z' sampled from the prior distribution of the standard population, and synchronously updating the weights of the discriminator and encoder at this time until the error value corresponding to the adversarial loss function converges; and Step S6.5: After training each batch, extract statistical features from the generated samples; then compare the statistical features with the standard demographic features to calculate the value of the population representativeness constraint loss function; through this comparison, the output value of the population representativeness constraint loss function will be updated; the training process will continue until the output value of the population representativeness loss function reaches a preset target value; finally, the generative adversarial autoencoder network model is optimized in this way to ensure that the generative adversarial autoencoder network model can fully reflect the representative characteristics of the population when generating samples; wherein the generated samples are the feature matrix X'.

4. The method for constructing a standard distribution model of healthy people based on generative adversarial autoencoders according to claim 1, characterized in that: The selected prior distribution function is a Gaussian distribution with a mean of 0 and a standard deviation of 1.

5. The method for constructing a standard distribution model of healthy people based on generative adversarial autoencoders according to claim 1 or 2, characterized in that: The electronic health record data includes: clinical measurement data; Among them, the clinical measurement data includes one of the following or a combination thereof: demographic data, physical examination data, laboratory examination data, and imaging examination data; the demographic data includes one of the following or a combination thereof: gender, age, height, weight, body mass index, and waist circumference; the physical examination data includes one of the following or a combination thereof: diastolic blood pressure and systolic blood pressure; the laboratory examination data includes one of the following or a combination thereof: blood examination data and urine examination data; the imaging examination data includes one of the following or a combination thereof: examination data in an X-ray examination report, examination data in a CT examination report, examination data in an MRI examination report, and examination data in an ultrasound examination report.

6. The method for constructing a standard distribution model of healthy people based on generative adversarial autoencoders according to claim 1, characterized in that: The expression of the reconstruction loss function is: In formula (1), x represents the input sample of the encoder, that is, the input data X; x′ represents the output sample of the decoder, that is, the feature matrix X′; N represents the number of samples in the data set; i represents the i-th input sample of the encoder; x i represents the data value of the i-th input sample of the encoder; x′ i represents the data value of the i-th output sample of the decoder; The expression of the adversarial loss function is: In formula (2), z represents a random variable sampled from the latent space representation sampled from the prior distribution of the standard population; p z represents the distribution of the random variable z; D(z) represents the probability value output by the discriminator; Represents the distribution p z Find the expectation of z obtained by sampling; z′ is the output obtained after inputting real data into the generator; q z is the distribution related to the generator; D(z′) is the discriminator’s judgment result on z′; The population representativeness constraint loss function includes measurement indicators at the feature level and the overall level; Among them, the representative score at the feature level is defined as It is used to represent the similarity of generated data on population feature k. The overall representative score is defined as gl, which is used to measure the comprehensive similarity of generated data on all features. The expression of the representative constraint loss function is as follows: In formula (3), EP trial represents the set of generated samples that satisfy all constraints; EP pool represents a set of samples that meet any single feature condition; if the generated data cannot meet the representativeness requirements of all features, a weighted feature-level constraint is introduced, and the expression is as follows: In formula (4), K is the total number of features; w k is the importance weight of feature k; is the representative score at the feature level. The total loss function combines reconstruction loss, adversarial loss and representative constraint loss. The total loss function is as follows: In formula (5), α and β are loss weight coefficients.

7. The method for constructing a standard distribution model of healthy people based on generative adversarial autoencoders according to claim 2, characterized in that: The step S2 specifically includes: Step S2.1: performing data cleaning on the first data set; Step S2.2: performing compressed sensing on the first data set after data cleaning; Step S2.3: normalizing the first data set after compressed sensing; Step S2.4: performing data enhancement on the normalized first data set; and / or The step S9 specifically includes: Step S9.1: performing data cleaning on the second data set; Step S9.2: performing compressed sensing on the second data set after data cleaning; Step S9.3: normalizing the second data set after compressive sensing; Step S9.4: Perform data enhancement on the normalized second data set.

Images