A method, device, system and storage medium for predicting fatty liver risk

Abstract

The present invention provides a method, device, system and storage medium for predicting fatty liver risk, which relates to the technical field of fatty liver risk detection, including: obtaining DNA sequencing results corresponding to stool samples of persons to be tested; obtaining relative abundance data of intestinal microbiota according to DNA sequencing results; inputting relative abundance data and phenotypic information characterizing life and physical conditions into a fatty liver risk prediction model based on deep belief network technology to obtain prediction results. In the present invention, human intestinal microbial communities are used as research objects, and based on intestinal microbiome and bioinformatics ideas, a method for predicting fatty liver risk using deep belief networks is provided, which clarifies the microbial differences between fatty liver patients and healthy people. By identifying the risk of fatty liver in advance, individuals can take appropriate lifestyle adjustments and treatment methods to reduce the risk of disease, and have the characteristics of low cost, non-invasive sampling, high stability, and high added value.

Description

Technical Field

[0001] The present invention relates to the technical field of fatty liver risk detection, and more specifically, to a fatty liver risk prediction method, device, system and storage medium. Background Art

[0002] Fatty liver disease is a liver disease characterized by the abnormal accumulation of fats, primarily triglycerides, in liver cells. This condition usually occurs when there is a disorder in the way the liver processes fats.

[0003] Fatty liver disease is broadly classified into two categories: (1) Non-alcoholic fatty liver disease (NAFLD): Fatty liver disease that is not caused by alcohol consumption. It is associated with obesity, type 2 diabetes, and other metabolic syndrome symptoms. (2) Alcoholic fatty liver disease (AFLD): Caused by long-term excessive drinking.

[0004] Fatty liver usually has no symptoms in the early stages. However, as the disease progresses, the following symptoms may appear: such as fatigue, discomfort or pain in the right upper abdomen, weight loss or jaundice (in severe cases). Therefore, timely risk prediction and examination before the onset of the disease is the key to avoiding further development of fatty liver.

[0005] Currently, the risk prediction methods for fatty liver mainly include the following: (1) Clinical evaluation: including medical history inquiry, physical examination, assessment of lifestyle (such as diet, exercise) and family history, etc. Clinical evaluation mainly relies on information provided by the patient and may be affected by subjectivity and recall bias; (2) Blood biochemical indicators: including liver function tests, insulin resistance tests and blood lipid level tests, etc. Blood indicators may be affected by many factors and may appear normal in the early stages of fatty liver. They cannot be used alone to diagnose fatty liver. (3) Imaging evaluation: such as ultrasound, CT or MRI scans. Among them, ultrasound is not sensitive enough for mild fatty liver, and its accuracy is affected by the operator's technique and the patient's body shape; while CT scans have radiation risks, are expensive, and have limited identification of mild fatty liver. MRI is expensive and is not suitable as a routine screening tool; (4) Genetic and biomarker research: studying specific genetic variants and blood biomarkers. The research field is relatively new, and many potential biomarkers and genetic factors have not been fully verified or widely used, and some biomarkers of early fatty liver cannot be reflected; (5) Lifestyle assessment: analyzing dietary habits, exercise levels, etc. This approach is subject to the accuracy and honesty of individual reports and may not be objective enough.

[0006] In summary, although there are multiple methods available to predict the risk of fatty liver, each method has its limitations, such as the high cost of imaging evaluation, the accuracy and complexity of biomarker measurement may be problematic, the prediction accuracy is not high, and it cannot provide personalized risk assessment. Summary of the invention

[0007] In view of this, in order to solve the above technical problems, the present invention provides a method for predicting fatty liver risk, comprising:

[0008] Obtaining DNA sequencing results corresponding to the stool samples of the persons to be tested;

[0009] Obtaining relative abundance data of the intestinal microbiota of the person to be tested according to the DNA sequencing results;

[0010] The relative abundance data, as well as the phenotypic information corresponding to the person to be tested that characterizes the living and physical conditions, are input into a trained fatty liver risk prediction model based on deep belief network technology to obtain a prediction result.

[0011] Preferably, before obtaining the DNA sequencing result corresponding to the fecal sample of the subject, the method further comprises:

[0012] Obtaining microbial DNA sequencing data corresponding to each stool sample in an experimental group, wherein the experimental group includes healthy subjects and fatty liver subjects;

[0013] Determining the abundance information corresponding to each of the stool samples according to the microbial DNA sequencing data;

[0014] Obtaining phenotypic information of a subject corresponding to each stool sample in the experimental group, and determining highly important features therein as a feature set according to the abundance information and the phenotypic information;

[0015] A deep belief network model is constructed according to the feature set, and the deep belief network model is trained using a test set and a training set constructed based on the experimental group to obtain the trained fatty liver risk prediction model.

[0016] Preferably, determining the features of high importance as a feature set according to the abundance information and the phenotype information comprises:

[0017] Merging the phenotypic information and the abundance information to obtain combined training data;

[0018] Constructing a decision tree model based on the combined training data;

[0019] Calculate the importance score of each feature according to the decision tree model;

[0020] The features are sorted based on the importance scores, and a preset number of features with high importance scores are extracted according to the sorting to form the feature set.

[0021] Preferably, calculating the importance score corresponding to each feature according to the decision tree model includes:

[0022] Divide the decision tree model into nodes according to the features, and calculate the initial impurity corresponding to each node in the decision tree model and the sub-node impurity of each sub-node after the division;

[0023] For the node division of the feature, calculating the impurity reduction corresponding to each node according to the initial impurity and the child node impurity;

[0024] For all partitions corresponding to each feature, a weighted average of the reduction in impurity is calculated, and the weighted average is used as the importance score representing the reduction in comprehensive impurity corresponding to each feature.

[0025] Preferably, in the node division for the feature, according to the initial impurity and the child node impurity, in calculating the impurity reduction corresponding to each node, the entropy index is used to calculate the initial impurity and the child node impurity, and the calculation formula is:

[0026]

[0027] Among them, I Entropy (t) represents the entropy value of node t; i represents the index of the category, ranging from 1 to c; c represents the number of categories, p i is the proportion of samples belonging to category i at the node.

[0028] Preferably, the impurity reduction corresponding to each node is obtained by calculating the weighted difference between the initial impurity and the child node impurity, and the child node includes a left child node and a right child node, and the left child node impurity and the right child node impurity are correspondingly provided;

[0029] The calculation formula for the reduction in impurity is:

[0030]

[0031] Where, Impurity Decrease(t) represents the reduction of the impurity of node t; Initia1 (t) represents the initial impurity of node t; N represents the total number of samples on node t; N left and N right I represents the number of samples on the left child node and the number of samples on the right child node respectively; left (t) and I right (t) represent the impurity of the left child node and the impurity of the right child node respectively.

[0032] Preferably, constructing a deep belief network model based on the feature set, and training the deep belief network model using a test set and a training set constructed based on the experimental group to obtain the trained fatty liver risk prediction model, including:

[0033] The deep belief network model is constructed by sequentially stacking multiple restricted Boltzmann machines;

[0034] Pre-training the deep belief network model layer by layer using contrastive divergence method;

[0035] The deep belief network model that has been pre-trained layer by layer is adjusted through a back propagation algorithm to obtain the trained fatty liver risk prediction model.

[0036] Preferably, the method of pre-training the deep belief network model layer by layer using the contrastive divergence method comprises:

[0037] The features in the feature set are taken as visible units and initialized as feature vectors of training data;

[0038] Updating hidden units in parallel according to the visible units, then reconstructing according to the hidden units, and updating the visible units in parallel;

[0039] The hidden units are updated in parallel using the reconstructed visible units, and the calculation methods used are:

[0040] p(h j =1|V)=σ(b j +∑ i V i w ij );

[0041] Among them, p(h j =1|V) represents that when the visible unit V is given, the hidden unit h i The probability of being activated; σ represents the sigmoid function; b j is the hidden unit h i The bias term of i represents the i-th element in the visible unit; w ij Represents the connection of the visible unit v i and the hidden unit h i The weight between

[0042] The calculation method used in the reconstruction based on the hidden unit and the parallel updating of the visible unit is:

[0043] p(v i =1|H)=σ(a i +∑ j hj w ij );

[0044] Among them, p(v i =1|H) represents that when the hidden unit H is given, the visible unit v i The probability of being activated; σ represents the sigmoid function; a i Represents the visible unit v i The bias term of j represents the jth element in the hidden unit; w ij Represents the connection of the visible unit v i and the hidden unit h i The weight between

[0045] According to the contrastive divergence method, the weights are updated by calculating the difference between the data distribution and the reconstructed distribution.

[0046] Preferably, the weights are updated by calculating the difference between the data distribution and the reconstructed distribution according to the contrastive divergence method, and the calculation method used is:

[0047] Δw ij ∝[(v i h j ) data -(v i h j ) reconstruction ];

[0048] Among them, Δw ij Represents the weight w ij The difference value of (v i h j ) data represents the data distribution value; (v i h j ) reconstruction Represents the reconstructed distribution value.

[0049] In addition, to solve the above problems, the present invention also provides a fatty liver risk prediction device, comprising:

[0050] An acquisition module is used to obtain the DNA sequencing results corresponding to the stool samples of the tested persons;

[0051] A calculation module, used to obtain the relative abundance data of the intestinal microbiota of the person to be tested according to the DNA sequencing result;

[0052] The prediction module is used to input the relative abundance data and the phenotypic information corresponding to the life and physical conditions of the person to be tested into a trained fatty liver risk prediction model based on deep belief network technology to obtain a prediction result.

[0053] The present invention provides a method for predicting fatty liver risk, comprising: obtaining DNA sequencing results corresponding to stool samples of a person to be tested; obtaining relative abundance data of the intestinal microbiome of the person to be tested according to the DNA sequencing results; inputting the relative abundance data and the phenotypic information corresponding to the person to be tested that characterizes the living and physical conditions into a trained fatty liver risk prediction model based on deep belief network technology to obtain a prediction result. In the present invention, taking the human intestinal microbial community as the research object, based on the intestinal microbiome and bioinformatics ideas, a method for predicting fatty liver risk using a deep belief network (DBN) is provided, which clarifies the microbial differences between fatty liver patients and healthy people. By identifying the risk of fatty liver in advance, individuals can take appropriate lifestyle adjustments and treatment methods to reduce the risk of disease, and have the characteristics of low cost, non-invasive sampling, high stability, and high added value. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic diagram of the structure of the hardware operating environment involved in the embodiment of the method for predicting fatty liver risk of the present invention;

[0055] Figure 2 This is a schematic diagram of the process of Example 1 of the method for predicting fatty liver risk of the present invention;

[0056] Figure 3 Schematic diagram of the process of steps S400-S700 in Example 2 of the method for predicting fatty liver risk of the present invention;

[0057] Figure 4 This is a schematic diagram of a detailed process of step S500 in Example 2 of the method for predicting fatty liver risk of the present invention;

[0058] Figure 5 This is a schematic diagram of a detailed process of step S600 in Example 3 of the method for predicting fatty liver risk of the present invention;

[0059] Figure 6 This is a schematic diagram of a detailed process of step S630 in Example 3 of the method for predicting fatty liver risk of the present invention;

[0060] Figure 7 This is a schematic diagram of the ranking of feature importance in Example 3 of the method for predicting fatty liver risk of the present invention;

[0061] Figure 8 This is a schematic diagram of a detailed process of step S700 in Example 4 of the method for predicting fatty liver risk of the present invention;

[0062] Fig. 9 This is a schematic diagram of a deep belief network in Example 4 of the method for predicting fatty liver risk of the present invention;

[0063] Fig.10 This is a schematic diagram of the RBM structure in Example 4 of the method for predicting fatty liver risk of the present invention;

[0064] Fig.11 This is a schematic diagram of the module connections of the fatty liver risk prediction device of the present invention.

[0065] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0066] The embodiments of the present invention are described in detail below, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions.

[0067] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0068] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0069] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0070] like Figure 1 , which is a schematic diagram of the structure of the hardware operating environment of the terminal involved in the embodiment of the present invention.

[0071] The fatty liver risk prediction system of the embodiment of the present invention can be a PC, or a mobile terminal device such as a smart phone, a tablet computer or a portable computer. The fatty liver risk prediction system may include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005 and a communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen, an input unit such as a keyboard, and a remote control. The optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory or a stable memory, such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001. Optionally, the fatty liver risk prediction system may also include an RF (Radio Frequency) circuit, an audio circuit, a WiFi module, and the like. In addition, the fatty liver risk prediction system can also be configured with other sensors such as gyroscopes, barometers, hygrometers, thermometers, infrared sensors, etc., which will not be described here.

[0072] Those skilled in the art will understand that Figure 1 The fatty liver risk prediction system shown in the figure does not constitute a limitation thereto, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. Figure 1 As shown, the memory 1005 as a computer-readable storage medium may include an operating system, a data interface control program, a network connection program, and a fatty liver risk prediction program.

[0073] In summary, the present invention takes the human intestinal microbial community as the research object, and based on the intestinal microbiome and bioinformatics ideas, provides a method for predicting the risk of fatty liver using a deep belief network (DBN), which clarifies the microbial differences between fatty liver patients and healthy people. By identifying the risk of fatty liver in advance, individuals can take appropriate lifestyle adjustments and treatment methods to reduce the risk of disease. It has the characteristics of low cost, non-invasive sampling, high stability, and high added value.

[0074] Embodiment 1:

[0075] Reference Figure 2 The embodiment of the present invention provides a method for predicting fatty liver risk, comprising:

[0076] Step S100, obtaining the DNA sequencing result corresponding to the stool sample of the person to be tested;

[0077] It should be noted that intestinal microorganisms are microbial communities living in the human intestine, including bacteria, fungi, viruses, etc. These microorganisms play an important role in human health, involving digestion, immune system regulation, metabolism and other aspects.

[0078] Despite the above associations, more research is still needed to gain a deeper understanding of the exact mechanisms and impacts between intestinal microorganisms and fatty liver, so it is important to choose an appropriate method to predict fatty liver risk.

[0079] This invention provides a new approach for individualized management of fatty liver disease based on risk prediction of the intestinal microbiome. By combining advanced technologies such as metagenomics and bioinformatics, microbial characteristics associated with fatty liver disease risk can be identified, thus providing a more precise strategy for early intervention and treatment.

[0080] As mentioned above, the sample targeted in the present invention is a stool sample of a person to be tested. The stool sample provided is usually collected by a medical professional, or collected by oneself with the assistance of a specific sampling tool.

[0081] Step S200, obtaining the relative abundance data of the intestinal microbiota of the person to be tested according to the DNA sequencing result;

[0082] As mentioned above, sample collection should be carried out according to standardized procedures to ensure the quality and consistency of the samples. The collected fecal samples are subjected to DNA extraction, sequencing and analysis to obtain the relative abundance of the intestinal microbiota.

[0083] Step S300, inputting the relative abundance data and the phenotypic information corresponding to the life and physical conditions of the person to be tested into a trained fatty liver risk prediction model based on deep belief network technology to obtain a prediction result.

[0084] As mentioned above, the data input into the fatty liver risk prediction model includes two aspects: (1) relative abundance data; (2) phenotypic information.

[0085] As mentioned above, phenotypic information refers to the observable characteristics or properties of an organism, which is the result of the interaction between genotype (genetic information) and the environment. In humans, phenotypic information includes physical characteristics, physiological functions, behavioral characteristics, etc., basically any observable biological characteristics.

[0086] Phenotypic information may include a variety of personal characteristics and lifestyle factors, including but not limited to the following:

[0087] Table 1. Phenotypic information table required to be provided by the test subjects

[0088]

[0089] The above phenotypic information is very important in medical and health research because it helps doctors and researchers understand an individual's health status and possible health risks, as well as the correspondence between health status, lifestyle habits and fatty liver conditions.

[0090] As mentioned above, Deep Belief Networks (DBNs) are a deep learning neural network structure used for unsupervised feature learning and classification problems. DBNs are generative models consisting of multiple hidden layers, each of which learns high-level representations of data in an unsupervised manner. These networks are composed of several layers of Restricted Boltzmann Machines (RBMs) or autoencoders stacked together.

[0091] The basic components of a deep belief network include: (1) Restricted Boltzmann Machine (RBM): Each layer of DBNs is usually an RBM. RBM is a randomly generated neural network that can learn the probability distribution of input data. It contains a visible layer and a hidden layer, and the nodes in the layer are not connected to each other, but only to the nodes in the other layer. (2) Autoencoder: Autoencoders are sometimes used as building blocks of DBNs. Autoencoders obtain effective representations of input data through unsupervised learning.

[0092] The method of using deep belief network (DBN) technology combined with DNA sequencing and phenotypic information to predict the risk of fatty liver has the following advantages:

[0093] (1) Comprehensive analysis: Multi-source data fusion: Combining genetic information (DNA sequencing data of the intestinal microbiota) and individual lifestyle and physical condition (phenotypic information) to provide a more comprehensive risk assessment.

[0094] More accurate risk prediction: By integrating multiple data types, an individual's risk of fatty liver disease can be predicted more accurately.

[0095] (2) Powerful capabilities of deep learning: Efficient feature extraction: DBN can effectively extract key features from complex data sets, especially in processing unstructured data (such as DNA sequences).

[0096] Good generalization ability: After training, deep belief networks are able to handle new, unseen data and make accurate predictions.

[0097] (3) Ability to model complex relationships: Capturing complex nonlinear relationships: DBN is particularly suitable for revealing complex and nonlinear relationships hidden in the data, which is very important for understanding the complex interactions between microbiota and fatty liver.

[0098] Better understanding of disease mechanisms: Analyzing the relationship between intestinal microbiota and lifestyle and fatty liver will help to gain a deeper understanding of the pathogenesis of fatty liver.

[0099] (4) Potential for personalized medicine: Personalized prediction: This approach can provide customized risk assessments for individuals, thereby facilitating personalized prevention and treatment strategies.

[0100] Early intervention: Accurate risk prediction helps to identify high-risk individuals at an early stage and implement early intervention measures.

[0101] (5) Data-driven decision support: Assist clinical decision-making: Provide strong data-based support to help doctors develop more effective prevention and treatment plans.

[0102] Epidemiological research: It helps in epidemiological research and understanding the distribution and influencing factors of fatty liver in different populations.

[0103] (7) Scalability: The model can be expanded: The DBN model can be easily expanded to include more data types and new research findings, thereby continuously improving the accuracy and reliability of its predictions.

[0104] Overall, the use of deep belief network technology combined with DNA sequencing and phenotypic information to predict the risk of fatty liver provides a comprehensive, efficient and potential method that can play an important role in personalized medicine and precision medicine.

[0105] In summary, the present invention takes the human intestinal microbial community as the research object, and based on the intestinal microbiome and bioinformatics ideas, provides a method for predicting the risk of fatty liver using a deep belief network (DBN), which clarifies the microbial differences between fatty liver patients and healthy people. By identifying the risk of fatty liver in advance, individuals can take appropriate lifestyle adjustments and treatment methods to reduce the risk of disease. It has the characteristics of low cost, non-invasive sampling, high stability, and high added value.

[0106] Embodiment 2:

[0107] Reference Figure 3 The second embodiment of the present invention provides a method for predicting fatty liver risk. Based on the above embodiment 1, the step S100, before obtaining the DNA sequencing result corresponding to the fecal sample of the subject, further includes:

[0108] Step S400, obtaining microbial DNA sequencing data corresponding to each stool sample in an experimental group; the experimental group includes healthy subjects and fatty liver subjects;

[0109] The above experimental group is used to construct and train the model through data. Therefore, it is necessary to include healthy subjects without fatty liver and fatty liver subjects diagnosed with fatty liver.

[0110] Before step S400, sample collection, DNA extraction and metagenomic sequencing are also required.

[0111] The first step is to collect stool samples from healthy subjects and subjects with fatty liver. The collection method includes: dividing the subjects into two groups, one group is healthy people (healthy subjects), and the other group is fatty liver people (fatty liver subjects). Use a disposable sampling tool to collect samples from fresh stool samples of the subjects. Avoid mixing samples with urine, water or other substances that may affect the quality of the samples. Collect enough samples for analysis. Insufficient sample volume will lead to failure of library construction. Store the samples in a low temperature environment, usually between 2°C and 8°C. Avoid exposing samples to high temperatures to prevent bacterial proliferation.

[0112] The second step is to extract the DNA of the microbiome: the samples collected in the first step are pre-treated by centrifugation, filter paper filtration or other sample preparation methods to remove solid particles and other impurities. Cell disruption is achieved by mechanical methods (such as oscillators), chemical methods (such as protease treatment) or heat treatment. Use a suitable DNA extraction kit or method to extract microbial DNA. Typically, this involves mixing the sample with the extraction reagent and then separating the DNA from other cellular components by centrifugation. The extracted DNA usually needs to be purified and washed to remove potential contaminants and reagent residues. The extracted DNA is then concentrated to a suitable concentration. This is usually done by spin concentration or other concentration methods. The extracted DNA is quality tested, usually using an ultraviolet spectrophotometer (UV-Vis spectrophotometer) or a fluorescence analyzer. This helps to determine the concentration and purity of the DNA.

[0113] The third step is metagenomic sequencing: PCR amplification of DNA is performed to increase the amount of DNA that can be sequenced. The extracted DNA fragments are constructed into a DNA library. This includes adapting the DNA fragments to the primers and connectors required by the sequencing platform. High-throughput sequencing technologies such as Illumina, 454, Ion Torrent, etc. are used to sequence the DNA library. These platforms are capable of sequencing a large number of DNA fragments at the same time, generating millions to billions of short sequences (short reads). Finally, the sequencing data is quality controlled, including removing low-quality sequences, removing adapter sequences, removing primer sequences, etc., to improve the accuracy of subsequent analysis, that is, obtaining the microbial DNA sequencing data.

[0114] Step S500, determining the abundance information corresponding to each of the stool samples according to the microbial DNA sequencing data;

[0115] For further reference, Figure 4 The step S500, determining the abundance information corresponding to each of the stool samples according to the microbial DNA sequencing data, includes:

[0116] Step S510, performing quality control on the microbial DNA sequencing data to obtain cleaned sequencing data;

[0117] The above-mentioned quality control methods include, but are not limited to: removing low-quality sequences, removing contaminating sequences, and repairing sequences.

[0118] As mentioned above, the microbial DNA sequencing data can be a sequence file in FASTQ format. In the quality control step, KneadData can be used for quality control, and the output will be the sequencing data that has been processed by quality control.

[0119] Step S520 , analyzing the cleaned sequencing data to determine the microbial species present therein and the abundance information corresponding to different classification levels of the microbial species.

[0120] The above-mentioned classification levels of microbial species include: domain, phylum, class, order, family, genus, and species.

[0121] As mentioned above, MetaPhlAn can be used to analyze the cleaned sequencing data. The sequencing data processed by KneadData is input into MetaPhlAn, which will analyze the microbial species present in the sample, estimate the relative abundance, and generate a species abundance table file.

[0122] This file contains all microbial species detected in each sample and their relative abundance at the domain, phylum, class, order, family, genus, and species levels, including the abundance information of the gut microbiota.

[0123] Step S600, obtaining phenotypic information of the subject corresponding to each stool sample in the experimental group, and determining highly important features therein as a feature set according to the abundance information and the phenotypic information;

[0124] Step S700, constructing a deep belief network model according to the feature set, and training the deep belief network model using a test set and a training set constructed based on the experimental group to obtain the trained fatty liver risk prediction model.

[0125] As mentioned above, phenotypic information refers to observable characteristics or properties of a subject, which is the result of the interaction between genotype (genetic information) and the environment. In humans, phenotypic information includes physical characteristics, physiological functions, behavioral characteristics, etc., and basically refers to any observable biological characteristics. For details, see Table 1 in Example 1.

[0126] As mentioned above, high importance may be a high importance score. The importance score of each feature may be obtained by calculation, and features corresponding to high importance scores may be screened out by sorting.

[0127] In this embodiment, not all phenotypic information and abundance information corresponding to microbial species are used as features for further judgment by the model. Instead, feature selection and screening are performed before building and training the model, and the importance scores of the features are determined. Highly important features are extracted as features in the feature set. This method has the following functions and advantages:

[0128] (1) Improve model performance, reduce the risk of overfitting, reduce the number of irrelevant or minor features, and help the model generalize better to new data. Enhance model accuracy. Using the most important features can improve the accuracy of model predictions.

[0129] (2) Improve computing efficiency and reduce computing burden. Fewer features means less data needs to be processed during model training and prediction, thereby reducing computing time and resource consumption. Accelerate model training. Reducing the number of features can speed up model training, especially when processing large data sets.

[0130] (3) Enhance model interpretability and improve result interpretability: Identifying and focusing on key features helps to better understand the model's predictions and underlying biological mechanisms. Facilitate result interpretation: In clinical and biomedical research, a simplified feature set makes it easier for doctors and researchers to interpret and use the model's results.

[0131] (4) Improve feature understanding and enhance understanding of data: By analyzing the importance of features, researchers can gain a deeper understanding of how various variables affect the development and progression of the disease. Reveal potential biomarkers: In medical research, these important features may become new biomarkers for the diagnosis and treatment of diseases.

[0132] (5) It helps in personalized medicine and provides a basis for personalized treatment: When predicting the risk of fatty liver disease, the identified key microorganisms and phenotypic characteristics (such as BMI) can be used to develop personalized prevention and treatment strategies.

[0133] (6) Promote research development and promote follow-up research: Determining which characteristics are most important can guide subsequent research work, especially in studies exploring the relationship between the microbiome and disease.

[0134] In summary, through feature selection and importance scoring, more efficient, accurate and easier to interpret models can be constructed, which is especially important when dealing with complex data sets such as intestinal microbiome data. This not only improves the practicality of the model, but also provides help for in-depth research and understanding of the biological basis of related diseases.

[0135] Embodiment 3:

[0136] Reference Figure 5 The third embodiment of the present invention provides a method for predicting fatty liver risk. Based on the above embodiment 2, the step S600 determines the features with high importance as a feature set according to the abundance information and the phenotype information, including:

[0137] Step S610, merging the phenotypic information and the abundance information to obtain merged training data;

[0138] In the above, the abundance information is the abundance information of the intestinal microbiome, and the phenotypic information and the abundance information are used together as features to be screened, and combined to obtain combined training data for further screening.

[0139] Step S620, constructing a decision tree model according to the combined training data;

[0140] In this embodiment, feature screening is performed by constructing a decision tree and using entropy as a measure of impurity.

[0141] The decision tree is built using the training data. In this process, the decision tree algorithm tries to find the best way to split the data so that the samples can be classified according to the different values ​​of the features.

[0142] Step S630, calculating the importance score corresponding to each feature according to the decision tree model;

[0143] Step S640, sorting the features based on the importance scores, and extracting a preset number of features with high importance scores according to the sorting to form the feature set.

[0144] In the above, the features are sorted according to the importance scores to obtain a sorted list, and the features with high importance scores that rank in the top preset number are extracted as the filtered features to form a feature set.

[0145] The preset number mentioned above can be determined according to the specific number of features ranked by importance score, for example, it can be set to 50, that is, the top 50 features with higher importance constitute the final feature set.

[0146] For further reference, Figure 6 The step S630, calculating the importance score corresponding to each feature according to the decision tree model, includes:

[0147] Step S631, dividing the decision tree model into nodes according to the features, calculating the initial impurity corresponding to each node in the decision tree model, and the sub-node impurity of each sub-node after the division;

[0148] The entropy metric is used to evaluate the reduction of node impurity. This metric measures the degree of mixing of samples on a node, and the reduction of node impurity indicates the amount by which the degree of sample mixing is reduced by splitting the node. Before splitting a node, entropy is first used to calculate the initial impurity of the node.

[0149] After the node is divided, the impurity of each child node is calculated. For each child node, the above calculation is repeated to obtain the child node impurity.

[0150] Specifically, in the node division for the feature, according to the initial impurity and the child node impurity, the impurity reduction corresponding to each node is calculated, and the entropy index is used to calculate the initial impurity and the child node impurity, and the calculation formula is:

[0151]

[0152] Among them, I Entropy (t) represents the entropy value of node t; i represents the index of the category, ranging from 1 to c; c represents the number of categories, p i is the proportion of samples belonging to category i at the node.

[0153] Step S632, for the node division of the feature, calculate the impurity reduction corresponding to each node according to the initial impurity and the child node impurity;

[0154] Next, for each feature split, the impurity reduction is calculated. The impurity reduction is the weighted difference between the initial impurity and the child node impurity.

[0155] Specifically, for the partition of node t, the reduction in impurity can be calculated using the following method:

[0156] In step S632, the impurity corresponding to each node is reduced by calculating the weighted difference between the initial impurity and the child node impurity, wherein the child node includes a left child node and a right child node, and the left child node impurity and the right child node impurity are correspondingly set;

[0157] The calculation formula for the reduction in impurity is:

[0158]

[0159] Where, Impurity Decrease(t) represents the reduction of the impurity of node t; Initial (t) represents the initial impurity of node t; N represents the total number of samples on node t; N left and N right I represents the number of samples on the left child node and the number of samples on the right child node respectively; left (t) and I right (t) represent the impurity of the left child node and the impurity of the right child node respectively.

[0160] Step S633: For all the partitions corresponding to each feature, a weighted average value of the reduction in impurity is calculated, and the weighted average value is used as the importance score representing the reduction in comprehensive impurity corresponding to each feature.

[0161] Finally, for all the partitions of each feature, calculate the weighted average of its impurity reduction. This will provide a comprehensive impurity reduction metric for each feature. Use the impurity reduction metric as the importance score of each feature, sort the features from high to low according to their importance scores, and select the top N most important features as the final feature set.

[0162] In this embodiment, taking the preset number = 50 as an example, the first 50 features with high importance are taken, such as Figure 7As shown in the figure. As can be seen from the figure, the most important intestinal bacteria in the risk prediction of fatty liver are Lachnospiraceae, Bacteroides, Escherichia, etc. Lachnospiraceae is a type of Gram-positive bacteria, which is usually enriched in the intestine and participates in the fermentation of food and the production of metabolites such as short-chain fatty acids (SCFA). Bacteroides is a type of Gram-negative bacteria that has a certain ability to decompose and absorb polysaccharides and is also involved in immune regulation in the intestine. Escherichia is a Gram-negative intestinal bacterium with multiple strains, including some that are part of the normal intestinal flora, while others may be involved in inflammation and destruction of the intestinal mucosa, are associated with inflammatory diseases and metabolic disorders, and are potential pathogens. In addition to the intestinal flora, the most important phenotypic feature is BMI, which is an indicator used to measure weight relative to height and is often used to assess whether a person is in the range of normal weight, overweight or obesity. Obesity is considered a major risk factor for fatty liver. Therefore, the importance of BMI reflects the direct impact of weight management on the risk of fatty liver.

[0163] Embodiment 4:

[0164] Reference Figure 8 The fourth embodiment of the present invention provides a method for predicting fatty liver risk. Based on the above embodiment 2, the step S700 constructs a deep belief network model according to the feature set, and trains the deep belief network model using a test set and a training set constructed based on the experimental group to obtain the trained fatty liver risk prediction model, including:

[0165] Step S710, constructing the deep belief network model by sequentially stacking multiple restricted Boltzmann machines;

[0166] The feature set selected in the previous step is used as the input feature, and the model will learn the mapping relationship from the input feature to the target variable (i.e., fatty liver risk). Fig. 9 As shown in Figure 1, a deep belief network consists of multiple layers of latent variables ("hidden units"), with connections between layers but no connections between units within each layer. The arrows between layers represent directed connections in the graphical model that the network represents.

[0167] As mentioned above, the deep belief network is constructed by stacking multiple restricted Boltzmann machines (RBM) in sequence. The schematic diagram of the RBM structure is as follows: Fig.10 shown.

[0168] As mentioned above, the restricted Boltzmann machine is a neural network used for unsupervised learning. It is a variant of the Boltzmann machine with one visible layer and one hidden layer, and the neurons within the layer are not connected to each other, only the neurons across the layers are connected. This structural feature enables RBM to effectively learn the feature representation of data.

[0169] Deep belief network models DBNs built by stacking RBMs provide a powerful mechanism in the field of deep learning, especially in feature learning and data representation. Although training DBNs can be computationally complex, their ability to understand deep data structures and features makes them very valuable in many applications.

[0170] Step S720, pre-training the deep belief network model layer by layer using contrastive divergence method;

[0171] As mentioned above, Contrastive Divergence (CD) is an algorithm for training restricted Boltzmann machines (RBMs) and deep belief networks (DBNs). CD is a fast and effective algorithm for approximating the gradients of training RBMs, especially in layer-by-layer pre-training of DBNs.

[0172] As mentioned above, the learning process of the deep belief network can be divided into the following two stages:

[0173] (1) Pre-train the RBM layer by layer first;

[0174] (2) Then use the backpropagation algorithm to fine-tune the network as a whole.

[0175] The RBM model was trained using the contrastive divergence (CD) method.

[0176] The contrastive divergence method used in this embodiment provides an approximation of the maximum likelihood method and is an ideal method for learning weights.

[0177] Step S730, adjusting the deep belief network model that has been pre-trained layer by layer through a back propagation algorithm to obtain the trained fatty liver risk prediction model.

[0178] As mentioned above, the backpropagation algorithm is a learning algorithm widely used in neural networks, which is used to effectively adjust the weights of each layer in the network according to the error of the network output.

[0179] Furthermore, the step S720, using contrastive divergence method to perform layer-by-layer pre-training on the deep belief network model, includes:

[0180] Step S721, taking the features in the feature set as visible units and initializing them as feature vectors of training data;

[0181] Step S722, updating the hidden units in parallel according to the visible units, and then reconstructing according to the hidden units and updating the visible units in parallel;

[0182] Step S723, using the reconstructed visible units to update the hidden units in parallel;

[0183] The calculation method used is: p(h j =1|V)=σ(b j +∑ i v i w ij );

[0184] Among them, p(h j =1|V) represents that when the visible unit V is given, the hidden unit h i The probability of being activated; σ represents the sigmoid function; b j is the hidden unit h i The bias term of i represents the i-th element in the visible unit; w ij Represents the connection of the visible unit v i and the hidden unit h i The weight between

[0185] Step S724: Update the weights by calculating the difference between the data distribution and the reconstructed distribution according to the contrastive divergence method.

[0186] Furthermore, in step S724, according to the contrastive divergence method, the weights are updated by calculating the difference between the data distribution and the reconstructed distribution, and the calculation method used is:

[0187] Δw ij ∝[(v i h j ) data -(v i h j ) reconstruction ];

[0188] Among them, Δw ij Represents the weight w ij The difference value of (v i h j ) data represents the data distribution value; (v i h j ) reconstruction Represents the reconstructed distribution value.

[0189] Above, contrastive divergence is applied to each subnetwork in turn, starting with the lowest layer (the lowest visible layer is a training set). Once one RBM is trained, another RBM is "stacked" on top of it and takes input from the last trained layer. The new visible layer is initialized to the training vector and the units in the trained layer are assigned values ​​using the current weights and biases. The new RBM is then trained as described above. The whole process is repeated until the desired stopping criterion is reached.

[0190] Once all layers have been pre-trained layer by layer, the deep belief network will enter the overall fine-tuning stage: the label data is introduced in the overall fine-tuning stage, that is, the clinical data of patients with fatty liver and healthy people. Using the label data, each layer of the deep belief network will further fine-tune its weights and biases. This is usually done using the back-propagation algorithm. The back-propagation algorithm adjusts the network parameters according to the principle of gradient descent to minimize the loss function (cross entropy). The entire deep belief network is optimized through multiple iterations of overall fine-tuning to improve its performance. This process needs to be repeated many times until the performance of the network converges to a satisfactory level.

[0191] Finally, the trained model can be used to predict the samples in the test set. The model will predict the risk of fatty liver for each test sample based on the input feature values.

[0192] In addition, the trained model can be validated and evaluated, and the performance indicators of the model can be calculated by comparing the model's prediction results with the actual labels in the test set (the true value of fatty liver risk).

[0193] Specifically, the accuracy is selected as the performance indicator of the model, that is, the ratio of the number of correctly predicted samples to the total number of samples.

[0194] Through grid search, the optimal hyperparameter hidden layer structure was determined to be [256, 256]. At this time, the model was optimal with an accuracy of 0.825. Compared with the random forest, the accuracy was improved by 0.025, indicating that the model has a good effect in predicting fatty liver risk.

[0195] In addition, reference Fig.11 This embodiment also provides a fatty liver risk prediction device, comprising:

[0196] An acquisition module 10 is used to obtain a DNA sequencing result corresponding to a stool sample of a person to be tested;

[0197] A calculation module 20, used to obtain the relative abundance data of the intestinal microbiota of the person to be tested according to the DNA sequencing result;

[0198] The prediction module 30 is used to input the relative abundance data and the phenotypic information corresponding to the life and physical conditions of the person to be tested into a trained fatty liver risk prediction model based on deep belief network technology to obtain a prediction result.

[0199] In addition, this embodiment also provides a fatty liver risk prediction system, including a memory and a processor, wherein the memory stores a fatty liver risk prediction program, and the processor runs the fatty liver risk prediction program to enable the fatty liver risk prediction system to execute the fatty liver risk prediction method described above.

[0200] In addition, this embodiment also provides a computer-readable storage medium, on which a fatty liver risk prediction program is stored. When the fatty liver risk prediction program is executed by a processor, the fatty liver risk prediction method as described above is implemented.

[0201] In summary, the present invention takes the human intestinal microbial community as the research object, and based on the intestinal microbiome and bioinformatics ideas, provides a method for predicting the risk of fatty liver using a deep belief network (DBN), which clarifies the microbial differences between fatty liver patients and healthy people. By identifying the risk of fatty liver in advance, individuals can take appropriate lifestyle adjustments and treatment methods to reduce the risk of disease. It has the characteristics of low cost, non-invasive sampling, high stability, and high added value.

[0202] Specifically, the fatty liver risk prediction method can have the following beneficial effects:

[0203] (1) Non-invasive and low-cost: Compared with traditional fatty liver risk assessment methods (such as blood pressure measurement and biochemical tests), this method uses stool samples for microbiota analysis, which is a non-invasive and low-cost method that does not require the collection of blood or other body fluid samples.

[0204] (2) Comprehensive application of microbial data: Using microbial data as input to the visible layer allows DBN to use this data to predict fatty liver risk. This takes advantage of the known association between gut microbes and health.

[0205] (3) Feature selection: Feature selection can be used to identify features that contribute less to the target variable, thereby reducing the input dimension. This helps reduce the complexity of the DBN model, improve training efficiency, and reduce the risk of overfitting.

[0206] (4) Hierarchical information extraction: The deep hierarchical structure of DBN allows the extraction and learning of features at different levels layer by layer. This is very useful when analyzing complex data.

[0207] (5) Improved interpretability: Using feature importance ranking for feature selection can make the final deep belief network easier to understand, especially when it is necessary to explain and communicate model results.

[0208] (6) Increased prediction accuracy: The accuracy of fatty liver risk prediction can be improved by combining microbiome data with other clinical parameters. This comprehensive analysis of multimodal data can provide a more comprehensive understanding of the potential risk factors for fatty liver.

[0209] (7) Personalized prediction: DBN based on microbial data can generate personalized risk assessments because each person’s microbiome is unique. This helps better meet the individual health needs of patients.

[0210] (8) Improved early intervention: Earlier identification of fatty liver risk can help doctors and patients take early intervention measures and improve the effectiveness of treatment.

[0211] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0212] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product is stored in a storage medium (such as ROM / RAM, disk, CD) as described above, including a number of instructions for a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention. The above is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly used in other related technical fields, are similarly included in the patent protection scope of the present invention.

Claims

1. A method for predicting fatty liver risk, characterized in that: include: Obtaining DNA sequencing results corresponding to the stool samples of the persons to be tested; Obtaining relative abundance data of the intestinal microbiota of the person to be tested according to the DNA sequencing results; Inputting the relative abundance data and the phenotypic information corresponding to the person to be tested that characterizes the life and physical condition into a trained fatty liver risk prediction model based on deep belief network technology to obtain a prediction result; Before obtaining the DNA sequencing result corresponding to the stool sample of the person to be tested, the method further includes: Obtaining microbial DNA sequencing data corresponding to each stool sample in an experimental group, wherein the experimental group includes healthy subjects and subjects with fatty liver; Determining the abundance information corresponding to each of the stool samples according to the microbial DNA sequencing data; Obtaining phenotypic information of a subject corresponding to each stool sample in the experimental group, and determining highly important features therein as a feature set according to the abundance information and the phenotypic information; According to the feature set, a deep belief network model is constructed, and the deep belief network model is trained using a test set and a training set constructed based on the experimental group to obtain the trained fatty liver risk prediction model; The step of determining, according to the abundance information and the phenotype information, features of high importance as a feature set comprises: Merging the phenotypic information and the abundance information to obtain combined training data; Constructing a decision tree model based on the combined training data; Calculate the importance score of each feature according to the decision tree model; Sorting the features based on the importance scores, and extracting a preset number of features with high importance scores according to the order to form the feature set; The importance score corresponding to each feature is calculated according to the decision tree model, including: Divide the decision tree model into nodes according to the features, and calculate the initial impurity corresponding to each node in the decision tree model and the sub-node impurity of each sub-node after the division; For the node division of the feature, calculating the impurity reduction corresponding to each node according to the initial impurity and the child node impurity; For all partitions corresponding to each feature, calculate the weighted average of the reduction in impurity, and use the weighted average as the importance score representing the reduction in comprehensive impurity corresponding to each feature; The step of constructing a deep belief network model based on the feature set, and training the deep belief network model using a test set and a training set constructed based on the experimental group to obtain the trained fatty liver risk prediction model includes: The deep belief network model is constructed by sequentially stacking multiple restricted Boltzmann machines; Pre-training the deep belief network model layer by layer using contrastive divergence method; The deep belief network model that has been pre-trained layer by layer is adjusted by a back propagation algorithm to obtain the trained fatty liver risk prediction model; The method of pre-training the deep belief network model layer by layer by using the contrastive divergence method includes: The features in the feature set are taken as visible units and initialized as feature vectors of training data; Updating hidden units in parallel according to the visible units, then reconstructing according to the hidden units, and updating the visible units in parallel; The hidden units are updated in parallel using the reconstructed visible units, and the calculation method used is: ； in, Represents the given visible unit When the hidden unit Probability of being activated; Represents the sigmoid function; is the hidden unit The bias term of Represents the i-th element in the visible unit; Represents the visible unit connected and the hidden units The weight between The calculation method used in the reconstruction based on the hidden unit and the parallel updating of the visible unit is: ； in, Represents that when the hidden unit H is given, the visible unit Probability of being activated; Represents the sigmoid function; Represents visible unit The bias term of represents the jth element in the hidden unit; Represents the visible unit connected and the hidden units The weight between According to the contrastive divergence method, the weights are updated by calculating the difference between the data distribution and the reconstructed distribution; According to the contrastive divergence method, the weights are updated by calculating the difference between the data distribution and the reconstructed distribution. The calculation method used is: ； in, Representative weight The difference value of Represents the data distribution value; Represents the reconstructed distribution value.

2. The method for predicting fatty liver risk according to claim 1, characterized in that: In the node division for the feature, according to the initial impurity and the child node impurity, the impurity reduction corresponding to each node is calculated, and the entropy index is used to calculate the initial impurity and the child node impurity, and the calculation formula is: ； in, represents the entropy value of node t; i represents the index of the category, ranging from 1 to c; c represents the number of categories, is the proportion of samples belonging to category i at the node.

3. The method for predicting fatty liver risk according to claim 1, characterized in that: The impurity reduction corresponding to each node is obtained by calculating the weighted difference between the initial impurity and the child node impurity, wherein the child node includes a left child node and a right child node, and the left child node impurity and the right child node impurity are correspondingly provided; The calculation formula for the reduction in impurity is: ； in, represents the reduction of impurity of node t; represents the initial impurity of node t; N represents the total number of samples on node t; and Respectively represent the number of samples on the left child node and the number of samples on the right child node; and Represent the impurity of the left child node and the impurity of the right child node respectively. .A device for predicting fatty liver risk, comprising: An acquisition module is used to obtain the DNA sequencing results corresponding to the stool samples of the tested persons; A calculation module, used to obtain the relative abundance data of the intestinal microbiota of the person to be tested according to the DNA sequencing result; A prediction module, used to input the relative abundance data and the phenotypic information corresponding to the tested person characterizing the living and physical conditions into a trained fatty liver risk prediction model based on deep belief network technology to obtain a prediction result; Before obtaining the DNA sequencing result corresponding to the stool sample of the person to be tested, the method further includes: Obtaining microbial DNA sequencing data corresponding to each stool sample in an experimental group, wherein the experimental group includes healthy subjects and subjects with fatty liver; Determining the abundance information corresponding to each of the stool samples according to the microbial DNA sequencing data; Obtaining phenotypic information of a subject corresponding to each stool sample in the experimental group, and determining highly important features therein as a feature set according to the abundance information and the phenotypic information; According to the feature set, a deep belief network model is constructed, and the deep belief network model is trained using a test set and a training set constructed based on the experimental group to obtain the trained fatty liver risk prediction model; The step of determining, according to the abundance information and the phenotype information, features of high importance as a feature set comprises: Merging the phenotypic information and the abundance information to obtain combined training data; Constructing a decision tree model based on the combined training data; Calculate the importance score of each feature according to the decision tree model; Sorting the features based on the importance scores, and extracting a preset number of features with high importance scores according to the order to form the feature set; The importance score corresponding to each feature is calculated according to the decision tree model, including: Divide the decision tree model into nodes according to the features, and calculate the initial impurity corresponding to each node in the decision tree model and the sub-node impurity of each sub-node after the division; For the node division of the feature, calculating the impurity reduction corresponding to each node according to the initial impurity and the child node impurity; For all partitions corresponding to each feature, calculate the weighted average of the reduction in impurity, and use the weighted average as the importance score representing the reduction in comprehensive impurity corresponding to each feature; The step of constructing a deep belief network model based on the feature set, and training the deep belief network model using a test set and a training set constructed based on the experimental group to obtain the trained fatty liver risk prediction model includes: The deep belief network model is constructed by sequentially stacking multiple restricted Boltzmann machines; Pre-training the deep belief network model layer by layer using contrastive divergence method; The deep belief network model that has been pre-trained layer by layer is adjusted by a back propagation algorithm to obtain the trained fatty liver risk prediction model; The method of pre-training the deep belief network model layer by layer by using the contrastive divergence method includes: The features in the feature set are taken as visible units and initialized as feature vectors of training data; Updating hidden units in parallel according to the visible units, then reconstructing according to the hidden units, and updating the visible units in parallel; The hidden units are updated in parallel using the reconstructed visible units, and the calculation method used is: ； in, Represents the given visible unit When the hidden unit Probability of being activated; Represents the sigmoid function; is the hidden unit The bias term of Represents the i-th element in the visible unit; Represents the visible unit connected and the hidden units The weight between The calculation method used in the reconstruction based on the hidden unit and the parallel updating of the visible unit is: ； in, Represents that when the hidden unit H is given, the visible unit Probability of being activated; Represents the sigmoid function; Represents visible unit The bias term of represents the jth element in the hidden unit; Represents the visible unit connected and the hidden units The weight between According to the contrastive divergence method, the weights are updated by calculating the difference between the data distribution and the reconstructed distribution; According to the contrastive divergence method, the weights are updated by calculating the difference between the data distribution and the reconstructed distribution. The calculation method used is: ； in, Representative weight The difference value of Represents the data distribution value; Represents the reconstructed distribution value.

Images