A method for constructing a depression risk prediction model
Through multimodal data algorithm and big data artificial intelligence technology, a depression risk prediction model is constructed from the social, psychological and biological levels, solving the problems of existing models in insufficient sample size and data complexity, and realizing the More accurate risk prediction and early diagnosis to adapt to real-time processing needs.
Patent Information
- Application Number
- CN202410889126.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-04
- Publication Date
- 2024-11-05
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing depression risk prediction model has problems with insufficient accuracy and generalization ability in case group sample size and complexity of data structures, especially in real-time processing scenarios.
Using multimodal data algorithms, combining big data and artificial intelligence technology, data is obtained from multiple levels such as society, psychology, and biology, and using deep learning and multiple machine learning algorithms to build a depression risk prediction model, including data preprocessing and characteristics Extraction, model training and cross-validation improve the generalization ability and computing efficiency of the model.
It achieves more accurate and reliable prediction of depression risk, improves early identification and diagnosis accuracy, reduces disability rates, and adapts to real-time data flow processing needs, with high computing efficiency.
Smart Images

Figure CN118430820B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and in particular to a method for constructing a depression risk prediction model. Background Art
[0002] Due to the high heterogeneity of depression, complex causes, and the influence of multiple factors such as social, psychological, and biological factors, the cause and pathogenesis of depression are still unclear to date. There is still a lack of an effective comprehensive system for early identification, objective diagnosis, and precise treatment of depression, and evidence-based evidence for treatment is still very insufficient.
[0003] The Chinese patent application number CN202311601376.3 discloses a method for constructing a risk prediction and assessment model for depression complicated with metabolic syndrome, including the following steps: obtaining MDD patient data from an electronic medical record database, screening out MDD patients who meet the criteria; obtaining covariate data of MDD patients, calculating the body mass index of each MDD patient through the covariate data, and evaluating metabolic syndrome in combination with blood biochemical indicators; screening out MDD patients with metabolic syndrome as the case group, and MDD patients without metabolic syndrome as the control group, obtaining the data of the case group patients as the case group data, and obtaining the data of the control group patients as the control group data; performing binary logistic regression analysis on the case group and control group data to obtain the results of relevant influencing factors; matching confounding factors using propensity scores, and using Bayesian network analysis to establish a risk prediction and assessment model according to the steps. The risk prediction and assessment model can accurately evaluate the risk of patients with depression with different characteristics combined with metabolic syndrome.
[0004] The above technical solution still has the following defects: ① Although it can accurately evaluate the risk of patients with different characteristics of depression combined with metabolic syndrome, in clinical practice, patients with depression combined with metabolic syndrome are rare, so there is an insufficient sample size in the case group, and the risk prediction assessment model constructed by the Bayesian network may not be accurate enough; ② Bayesian network is a machine learning algorithm that combines directed acyclic graphs and Bayesian probability, which can intuitively represent the joint probability distribution of multiple variables, process incomplete data, and estimate uncertainty through probabilistic inference. However, there are some challenges and limitations in using Bayesian networks for risk prediction, such as the Bayesian network inference algorithm is suitable for processing a single type of data, and requires sufficient sample data to train the model. The construction and inference results of the model may be limited by the data structure and prior knowledge, and the generalization ability may be relatively weak; structural learning is more difficult, and usually requires manual construction of the network structure or reliance on expert knowledge; the computational complexity of the inference algorithm is high and is not suitable for real-time processing scenarios.
[0005] Multimodal data algorithms have the following advantages when building risk prediction models: ① They can use data from different data sources or different types to provide more comprehensive information; ② They can fuse data features of different modalities to build more powerful and more representative feature representations, which helps to improve the prediction accuracy of the model; ③ They can learn the statistical laws and potential relationships of data from multiple data sources, thereby improving the generalization ability of the model; ④ They use flexible modeling methods such as deep learning to automatically learn complex relationships between data and adapt to different types of data and problems; ⑤ They can better adapt to the processing needs of real-time data streams, can process large-scale and high-dimensional data, and have high computational efficiency.
[0006] Therefore, building a depression risk prediction model and realizing a comprehensive system for early identification, objective diagnosis and precise treatment of depression is crucial to effectively prevent the occurrence of depression and assist clinicians in providing targeted and individualized guidance for the treatment of depression. Building a depression risk prediction model using multimodal data algorithms can more comprehensively utilize different types of data information and build a more accurate and reliable risk prediction model. Summary of the invention
[0007] The purpose of the present invention is to solve the above problems and provide a method for constructing a depression risk prediction model.
[0008] In order to achieve the above object, the technical solution of the present invention is as follows:
[0009] A method for constructing a depression risk prediction model comprises the following steps:
[0010] S1. Screening of patients with depression
[0011] Based on community / school cohorts and hospital psychiatric outpatient and ward platforms, patients with depression who met the criteria were screened as the case group, and healthy people were selected as the control group;
[0012] S2. Data Collection
[0013] Collect macro-micro environmental information data of the case group and the control group at baseline and during follow-up;
[0014] S3. Data Analysis
[0015] Perform statistical analysis on the case group and control group data collected in step S2, and obtain the results of relevant influencing factors of depression from the multiple levels of gene composition-transcriptional regulation-protein expression-intestinal microorganisms-metabolic pathways and brain functional network-cognition-behavior-environment;
[0016] S4. Early individualized warning
[0017] Realize early individualized warning for depression patients based on big data and artificial intelligence methods;
[0018] S5. Constructing risk prediction model
[0019] Propensity scores were used to match confounding factors, and a risk prediction model was constructed in steps using multimodal data algorithms.
[0020] Furthermore, in step S1, the method for screening depression patients is based on the community / school cohort and hospital psychiatric clinic and ward platform to build a real-time online survey platform, and conduct data verification, supplementation and improvement in the form of household surveys to form a survey of the large database of different life cycles of the entire population of the cluster sampling community. According to the diagnostic criteria of the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders, depression patients who meet the criteria are screened as the case group, and healthy people are selected as the control group.
[0021] Further, in step S2, the macro-micro environment information data includes but is not limited to social factors, physical environment, psychological environment, genetic factors, microbial environment, biochemistry and metabolism, and EEG-brain neuroimaging;
[0022] The social factors include, but are not limited to, economic and health care factors;
[0023] The psychological environment includes but is not limited to psychosocial, clinical characteristics and neuropsychological indicators.
[0024] Furthermore, in step S3, the method for performing statistical analysis on the collected case group and control group data is a binary logistic regression analysis method.
[0025] Further, in step S4, the specific operation of the early individualized warning is as follows:
[0026] S41: Systematically collect, track and integrate macro and micro variables of the general population and special populations into a large database of clinical research on brain diseases;
[0027] S42. Based on big data and artificial intelligence methods, individualized early warning for patients with depression can be achieved through advanced algorithms.
[0028] Further, in step S42, the specific operation of implementing individualized early warning for depression patients through advanced algorithms based on big data and artificial intelligence methods is as follows:
[0029] S421, first use the logistic regression model to perform prediction and classification;
[0030] S422. Compare and judge the prediction result of the model with the threshold. If the predicted probability of depression of the patient exceeds the threshold, issue a warning prompt.
[0031] Further, in step S421, the prediction function is:
[0032] p(y i =1|x i )=σ(w T x i +b),
[0033] Where p(y i =1|x i ) represents the probability of depression in patient i, y i The label of patient i, indicating whether the patient has depression or not; x i represents the feature vector of patient i, which contains various features or indicators used to predict depression, w is the weight vector of the model, b is the bias, σ(.) is the logistic function (logistic sigmoid), and T represents the transpose of the weight vector, the purpose of which is to calculate the weight of the feature vector.
[0034] Further, in step S422, the prediction result of the model is compared with the threshold value, and the judgment formula is as follows: ifp(y i =1|x i )>F, an early warning is issued, where p(y i =1|x i ) represents the probability of depression of patient i, and F represents the threshold.
[0035] Further, in step S5, the specific operations of constructing the risk prediction model are as follows: using the chi-square test to compare the differences in covariate data such as gender and age of patients in the case group and the control group, using propensity score matching to reduce the selection bias of the two groups, and using a multimodal data algorithm to establish a risk prediction assessment model;
[0036] All patients in the case group and the control group were matched 1:1 by age and gender.
[0037] Further, in step S5, the multimodal model algorithm includes the following steps:
[0038] S51. Multimodal data feature preprocessing: including data cleaning, scoring and conversion, outlier processing, and data normalization;
[0039] S52, data alignment: perform spatial alignment, temporal synchronization and inter-modal fusion on the preprocessed data in sequence.
[0040] S53, Deep learning and pattern recognition;
[0041] S54, cross validation: The 5-fold cross validation method is used to divide the data set into five subsets of equal size. Each time, one subset is used as a test set, and the remaining subsets are used to train the model. This is repeated five times to ensure that each subset has a chance to be used as a test set.
[0042] S55, Performance evaluation: A set of evaluation methods that follow strict scientific principles and statistical standards are adopted to comprehensively and objectively evaluate the performance of the developed algorithms in actual application scenarios and ensure the repeatability and reliability of the results.
[0043] Further, in step S53, the deep learning and pattern recognition includes:
[0044] ① Feature extraction: Use statistical methods such as weighted combination, principal component analysis (PCA), factor analysis, etc. to reduce the dimension of data while retaining the most valuable information, and use deep learning to extract useful key information from complex data;
[0045] ② Model training: Use a variety of machine learning algorithms to build different prediction models, and select the optimal classification model by comparing the prediction performance, so as to achieve high-precision prediction of depression risk;
[0046] ③ Convolutional Neural Network (CNN): It automatically and effectively extracts image features through convolutional layers, and gradually abstracts them into higher-level image semantic information through hierarchical structures. It uses sigmoid activation functions to implement binary classification in CNN, and automatically and accurately makes classification decisions by learning complex features of images.
[0047] ④ Residual Network (ResNet): By introducing the residual learning framework, the problem of deep network training difficulties is solved; ResNet can greatly improve the efficiency and accuracy of training. A fully connected layer is added to the end of the ResNet network. Using the sigmoid activation function, the network can output a probability value between 0 and 1, indicating the possibility that the sample belongs to a certain category;
[0048] ⑤ Attention Mechanism: The global dependency between elements in the input sequence is captured through the self-attention mechanism. In the end, the focused feature information is usually mapped to the binary output through one or more fully connected layers. The sigmoid activation function is usually used at the end of these fully connected layers to output a probability value between 0 and 1, indicating the possibility that the sample belongs to the target category.
[0049] ⑥Generative Adversarial Network (GAN): GAN consists of a generative network and a discriminative network, and the quality of the generated image is improved through the adversarial process between the two.
[0050] The use of artificial intelligence models in face image processing not only improves the accuracy of recognition and analysis, but also provides a powerful tool for in-depth understanding of the complex information in face images. The implementation of binary classification within the GAN framework mainly depends on the function of the discriminator D. The discriminator part of GAN can use the complex and deep image features it has learned to make high-accuracy classification decisions.
[0051] ⑦Self-attention model (Transformer): is a network model that performs powerfully on multiple visual tasks, including image classification, object detection and segmentation.
[0052] Furthermore, the core formula of the residual network is:
[0053] y=F(x,{W i})+x,
[0054] Where x and y are the input and output of the network respectively, F(x,{W i}) represents the learned residual mapping, W i Represents a set of weight parameters of a model, usually containing multiple weight vectors or matrices;
[0055] Furthermore, the core formula of the adversarial network is:
[0056]
[0057] Among them, G represents the generator, D represents the discriminator, minG represents that the generator G tries to minimize the objective function, maxD represents that the discriminator D tries to maximize the objective function, V(D,G) represents the value function of GAN, which represents the adversarial goal between the generator and the discriminator; E represents the expected value operator, x~p data (x) indicates that sample x comes from the real data distribution Pdata(x); logD(x) indicates the output of the discriminator D for the real sample x. The goal of the discriminator D is to judge the real data x as true (the output is close to 1), so the larger the value of logD(x), the better; z~p z (z) indicates that the input factor z comes from the factor distribution p z (z), G(z) represents the false data generated by the generator G according to the factor z, log(1-D(G(z))) represents the output of the discriminator D on the generated data G(z), and the goal of the discriminator D is to judge the generated data G(z) as false (the output is close to 0), so the larger the value of log(1-D(G(z))), the better.
[0058] In addition, it should be noted that the self-attention model (Transformer) involved in the present invention includes two components: an adaptive N×N convolutional encoder and a split deep transposed attention encoder.
[0059] First, the model extracts hierarchical features through four stages of different scales. The input image passes through a patchify stem layer, uses a 4×4 non-overlapping convolution, and is followed by a layer norm layer to generate a H / 4×W / 4×C1 feature map. Next, the output is passed to a 3×3 convolution layer, and the encoder extracts local features.
[0060] The second stage starts with a downsampling layer implemented with a 2×2 strided convolution, halving the spatial size and increasing the channels, generating a H / 8×W / 8×C2 feature map, followed by two consecutive 5×5 convolutional encoders. The position encoding is added only before the second stage. For dense prediction tasks such as object detection and segmentation, the patchify stem is very sensitive, so it is only added once in the network to encode the spatial position information. The output feature map continues to be passed to the third and fourth stages, generating H / 16×W / 16×C3 and H / 32×W / 32×C4 dimensional features, respectively.
[0061] A variety of performance evaluation indicators are used to conduct a comprehensive comparison with convolutional networks, ViT and other hybrid models to fully demonstrate the competitiveness of the model on various tasks.
[0062] The Transformer model implements binary classification through the final fully connected layer, which maps the high-dimensional features learned by the model through the self-attention mechanism and multi-layer perceptron to two categories.
[0063] In summary, the present invention has the following beneficial effects:
[0064] 1. The present invention establishes a multidimensional indicator system and model for depression risk prediction, individualized prevention, early diagnosis and precise treatment for the entire population and throughout the life cycle from multiple dimensions, multiple levels and their interactions such as society, psychology, biology (genetic sensitivity, neural circuits, intestinal flora, etc.).
[0065] 2. The present invention adopts artificial intelligence fusion analysis technology to monitor and analyze in real time the transformation from health to disease and from disease to recovery state in the natural state, as well as the influencing factors of medical intervention on disease outcome, to establish and develop early warning and risk prediction models for high-risk groups of depression, and to optimize the prediction model in real time through dynamic follow-up, ultimately improving the early risk prediction and diagnosis of depression, the early treatment rate, and reducing the disability rate of depression.
[0066] 3. The risk prediction model constructed by the present invention using a multimodal data algorithm can more comprehensively utilize different types of data information, fuse data features of different modalities, construct a more powerful and more representative feature representation, and improve the accuracy and reliability of model predictions; it can also better adapt to the processing needs of real-time data streams, can process large-scale and high-dimensional data, and has high computing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0067] Figure 1 It is a general flow chart of a method for constructing a depression risk prediction model of the present invention; DETAILED DESCRIPTION
[0068] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0069] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below in conjunction with the embodiments.
[0070] Example:
[0071] Depend on Figure 1 As shown, a method for constructing a depression risk prediction model in this embodiment includes the following steps:
[0072] S1. Screening of patients with depression
[0073] Based on the community / school cohort and the psychiatric outpatient and ward platforms of the Second Xiangya Hospital of Central South University, a real-time online survey platform was built, and combined with household surveys to verify, supplement and improve the data, forming a survey of the big database of the entire population in the cluster sampling community at different life cycles. According to the diagnostic criteria of the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), patients with depression who met the criteria were screened as the case group, and healthy people as the control group.
[0074] The specific diagnostic criteria for depression are: within 2 weeks, 5 or more of the 9 symptoms in the diagnostic criteria are met, and at least 1 of them is a core symptom: ① depressed mood; ② loss of interest or pleasure.
[0075] S2. Data Collection
[0076] The macro-micro environmental information data of the case group and the control group were collected at baseline and during follow-up; the macro-micro environmental information data included social factors (economy and medical services, etc.), physical environment, psychological environment (social psychology, clinical characteristics and neuropsychological indicators, etc.), genetic factors, microbial environment, biochemistry and metabolism, and EEG-brain neuroimaging, etc. The follow-up was 1 year, 2 years, and 3 years respectively.
[0077] Specifically, the types of macro-micro environmental factors are as follows:
[0078]
[0079]
[0080] S3. Data Analysis
[0081] The data of the case group and the control group collected in step S2 were statistically analyzed using binary logistic regression analysis methods to obtain the results of relevant influencing factors of depression from the multi-level levels of gene composition-transcriptional regulation-protein expression-intestinal microorganisms-metabolic pathways and brain functional networks-cognition-behavior-environment. Through data analysis, important influencing factors affecting depression were found, and some results are as follows:
[0082]
[0083] Current data show that a study of 500 patients with depression and 1,000 healthy people found that personality traits, adverse life events, family, school, social environment, brain gray matter volume, functional networks, partial EEG characteristics, intestinal flora, viruses, metabolites, inflammatory indicators, etc. are important influencing factors for patients with depression.
[0084] S4. Early individualized warning
[0085] Based on big data and artificial intelligence methods, early individualized warning for depression patients can be achieved; the specific operations are as follows:
[0086] S41: Systematically collect, track and integrate macroscopic, mesoscopic and microscopic variables of the general population and special populations into a large database of clinical research on brain diseases.
[0087] S42. Based on big data and artificial intelligence methods, advanced algorithms are used to achieve individualized early warning for patients with depression. The specific operations are as follows:
[0088] S421. First, a logistic regression model is used for prediction and classification. The prediction function is:
[0089] p(y i =1|x i )=σ(w T xi +b),
[0090] Where p(y i =1|x i ) represents the probability of depression in patient i, w is the weight vector of the model, b is the bias, σ(.) is the logistic function (logistic sigmoid), and T represents the transpose of the weight vector, the purpose of which is to calculate the weight of the feature vector.
[0091] S422. Compare and judge the prediction result of the model with the threshold. If the predicted probability of depression of the patient exceeds the threshold, issue a warning prompt. The judgment formula is as follows:
[0092] if p(y i =1|x i )>F, an early warning will be issued.
[0093] Where p(y i =1|x i ) represents the probability of depression of patient i, and F represents the threshold.
[0094] S5. Constructing risk prediction model
[0095] The chi-square test was used to compare the differences in covariate data such as gender and age between the case group and the control group. Propensity score matching was used to reduce the selection bias between the two groups. A risk prediction model was constructed using a multimodal model algorithm according to a 1:1 matching method.
[0096] Among them, the multimodal model algorithm includes the following steps:
[0097] S51. Multimodal data feature preprocessing: including data cleaning, scoring and conversion, outlier processing, and data normalization;
[0098] S52, data alignment: perform spatial alignment, temporal synchronization and inter-modal fusion on the preprocessed data in sequence.
[0099] S53. Deep learning and pattern recognition:
[0100] ① Feature extraction: Use statistical methods such as weighted combination, principal component analysis (PCA), factor analysis, etc. to reduce the dimension of data while retaining the most valuable information, and use deep learning to extract useful key information from complex data.
[0101] ② Model training: Use a variety of machine learning algorithms to build different prediction models, and select the optimal classification model by comparing the prediction performance, so as to achieve high-precision prediction of depression risk.
[0102] ③ Convolutional Neural Network (CNN): Automatically and effectively extract image features through convolutional layers, and gradually abstract them into higher-level image semantic information through a hierarchical structure. Use the sigmoid activation function to implement binary classification in CNN, and automatically and accurately make classification decisions by learning the complex features of the image.
[0103] ④ Residual Network (ResNet): It solves the problem of difficult deep network training by introducing the residual learning framework. Its core formula is:
[0104] y=F(x,{W i})+x,
[0105] Where x and y are the input and output of the network respectively, F(x,{W i}) represents the learned residual mapping, W i Represents a set of weight parameters of a model, usually containing multiple weight vectors or matrices.
[0106] ResNet can greatly improve the efficiency and accuracy of training. By adding a fully connected layer at the end of the ResNet network and using the sigmoid activation function, the network can output a probability value between 0 and 1, indicating the possibility that the sample belongs to a certain category.
[0107] ⑤ Attention Mechanism: The global dependencies between elements in the input sequence are captured through the self-attention mechanism, and the focused feature information is usually mapped to the binary output through one or more fully connected layers. The end of these fully connected layers usually uses a sigmoid activation function to output a probability value between 0 and 1, indicating the possibility that the sample belongs to the target category.
[0108] ⑥Generative Adversarial Network (GAN): GAN consists of a generative network and a discriminative network, and the quality of the generated image is improved through the adversarial process between the two.
[0109] The core formula is:
[0110]
[0111] Among them, G represents the generator, D represents the discriminator, minG represents that the generator G tries to minimize the objective function, maxD represents that the discriminator D tries to maximize the objective function, V(D,G) represents the value function of GAN, which represents the adversarial goal between the generator and the discriminator; E represents the expected value operator, x~p data(x) indicates that sample x comes from the real data distribution Pdata(x); logD(x) indicates the output of the discriminator D for the real sample x. The goal of the discriminator D is to judge the real data x as true (the output is close to 1), so the larger the value of logD(x), the better; z~p z (z) indicates that the input factor z comes from the factor distribution p z (z), G(z) represents the false data generated by the generator G according to the factor z, log(1-D(G(z))) represents the output of the discriminator D on the generated data G(z), and the goal of the discriminator D is to judge the generated data G(z) as false (the output is close to 0), so the larger the value of log(1-D(G(z))), the better.
[0112] The use of artificial intelligence models in face image processing not only improves the accuracy of recognition and analysis, but also provides a powerful tool for in-depth understanding of the complex information in face images. The implementation of binary classification within the GAN framework mainly depends on the function of the discriminator D. The discriminator part of GAN can use the complex and deep image features it has learned to make high-accuracy classification decisions.
[0113] ⑦Self-attention model (Transformer): is a network model that performs powerfully on multiple visual tasks, including image classification, object detection and segmentation.
[0114] The self-attention model (Transformer) involved in the present invention includes two components: an adaptive N×N convolutional encoder and a split deep transposed attention encoder.
[0115] First, the model extracts hierarchical features through four stages of different scales. The input image passes through a patchify stem layer, uses a 4×4 non-overlapping convolution, and is followed by a layer norm layer to generate a H / 4×W / 4×C1 feature map. Next, the output is passed to a 3×3 convolution layer, and the encoder extracts local features.
[0116] The second stage starts with a downsampling layer implemented with a 2×2 strided convolution, halving the spatial size and increasing the channels, generating a H / 8×W / 8×C2 feature map, followed by two consecutive 5×5 convolutional encoders. The position encoding is added only before the second stage. For dense prediction tasks such as object detection and segmentation, the patchify stem is very sensitive, so it is only added once in the network to encode the spatial position information. The output feature map continues to be passed to the third and fourth stages, generating H / 16×W / 16×C3 and H / 32×W / 32×C4 dimensional features, respectively.
[0117] A variety of performance evaluation indicators are used to conduct a comprehensive comparison with convolutional networks, ViT and other hybrid models to fully demonstrate the competitiveness of the model on various tasks.
[0118] The Transformer model implements binary classification through the final fully connected layer, which maps the high-dimensional features learned by the model through the self-attention mechanism and multi-layer perceptron to two categories.
[0119] S54, cross validation: The 5-fold cross validation method is used to divide the data set into five subsets of equal size. Each time, one subset is used as a test set, and the remaining subsets are used to train the model. This is repeated five times to ensure that each subset has a chance to be used as a test set.
[0120] S55, Performance evaluation: A set of evaluation methods that follow strict scientific principles and statistical standards are adopted to comprehensively and objectively evaluate the performance of the developed algorithms in actual application scenarios and ensure the repeatability and reliability of the results.
[0121] Comparative Example 1: Except that the model is constructed using a single-modal model algorithm, other steps are consistent with Example 1.
[0122] According to the evaluation of the accuracy of diagnosing depressive disorders based on questionnaire surveys and niacin skin phenotypes, the accuracy of diagnosing depressive disorders based on self-assessment screening of single-dimensional scales is 35.4%, the accuracy of diagnosing depressive disorders based on niacin phenotypes is 56.6%, and the accuracy of diagnosing depressive disorders by combining niacin phenotypes with multi-dimensional data of self-assessment scales is 68.1%.
[0123] Metabolomic analysis of depression and healthy controls revealed that the area under the AUC curve for pyruvate, malate, and linoleylcarnitine distinguished depression from healthy controls with an accuracy of approximately 90%.
[0124] Using resting-state EEG data, a random forest model was used to achieve classification of depression and healthy controls with an accuracy of 82.05%, a sensitivity of 84.62%, a specificity of 79.49% and an area under the curve (AUC) of 84.88%.
[0125] The present invention uses multimodal data algorithms to construct a multidimensional indicator system and model for risk prediction of depression, individualized prevention, early diagnosis and precision treatment for the entire population and the entire life cycle from multiple dimensions, multiple levels and their interactions such as society, psychology, biology (genetic sensitivity, neural circuits, intestinal flora, etc.). It can more comprehensively utilize different types of data information, integrate data features of different modalities, construct a more powerful and more representative feature representation, and improve the accuracy and reliability of model prediction; thereby achieving real-time monitoring and analysis of the transformation from health to disease and from disease to recovery under natural conditions, as well as the influencing factors of medical intervention on disease outcomes, and dynamically optimize the prediction model in real time through follow-up, ultimately improving the early risk prediction and diagnosis of depression, the early treatment rate, and reducing the disability rate of depression.
[0126] The above-described embodiments only express the preferred implementation modes of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present invention. It should be pointed out that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.
Claims
1. A method for constructing a depression risk prediction model, characterized in that: The following steps are involved: S1. Screening of patients with depression Based on community / school cohorts and hospital psychiatric outpatient and ward platforms, patients with depression who met the criteria were screened as the case group, and healthy people were selected as the control group; S2. Data Collection Collect macro-micro environmental information data of the case group and the control group at baseline and during follow-up; S3. Data Analysis Perform statistical analysis on the case group and control group data collected in step S2, and obtain the results of relevant influencing factors of depression from the multiple levels of gene composition-transcriptional regulation-protein expression-intestinal microorganisms-metabolic pathways and brain functional network-cognition-behavior-environment; S4. Early individualized warning Realize early individualized warning for depression patients based on big data and artificial intelligence methods; The specific operation of the early individualized warning is as follows: S41: Systematically collect, track and integrate macro and micro variables of the general population and special populations into a large database of clinical research on brain diseases; S42. Based on big data and artificial intelligence methods, advanced algorithms are used to achieve individualized early warning for patients with depression. The specific operations are as follows: S421. First, a logistic regression model is used for prediction and classification. The function predicted by the logistic regression model is: ; in represents the probability of depression for patient i, The label of patient i, which is used to indicate whether the patient has depression or not; represents the feature vector of patient i, which contains various features or indicators used to predict depression, w is the weight vector of the model, T represents the transpose of the weight vector, b is the bias, is a logical function; S422, comparing and judging the prediction result of the logistic regression model with the threshold, and issuing a warning prompt if the predicted probability of depression of the patient exceeds the threshold; S5. Constructing risk prediction model Use propensity scores to match confounding factors, utilize multimodal data algorithms, and follow the steps to build a risk prediction model; The multimodal data algorithm comprises the following steps: S51, multimodal data feature preprocessing; S52, data alignment; S53, deep learning and pattern recognition, the deep learning and pattern recognition including: ① feature extraction; ② model training; ③ convolutional neural network; ④ residual network; ⑤ attention mechanism; ⑥ generative adversarial network; ⑦ self-attention model; S54, cross validation; S55. Performance evaluation.
2. The method for constructing a depression risk prediction model according to claim 1, characterized in that: In step S1, the method for screening patients with depression is to build a real-time online survey platform based on community / school cohorts and hospital psychiatric clinics and ward platforms, and to conduct data verification, supplementation and improvement in the form of household surveys, so as to form a survey of the big database of different life cycles of the entire population of the cluster sampling community, and screen the patients with depression who meet the criteria as the case group and the healthy people as the control group.
3. The method for constructing a depression risk prediction model according to claim 1, characterized in that: In step S2, the macro-micro environmental information data includes but is not limited to social factors, physical environment, psychological environment, genetic factors, microbial environment, biochemistry and metabolism, and EEG-brain neuroimaging; The social factors include, but are not limited to, economic and health care factors; The psychological environment includes but is not limited to psychosocial, clinical characteristics and neuropsychological indicators.
4. The method for constructing a depression risk prediction model according to claim 1, characterized in that: In step S3, the method for performing statistical analysis on the case group and control group data collected in step S2 is a binary logistic regression analysis method.
5. The method for constructing a depression risk prediction model according to claim 1, characterized in that: In step S422, the prediction result of the logistic regression model is compared with the threshold value, and the judgment formula is as follows: A warning is issued, including represents the probability of depression of patient i, and F represents the threshold.
6. The method for constructing a depression risk prediction model according to claim 1, characterized in that: In step S5, the specific operations of constructing the risk prediction model are as follows: using the chi-square test to compare the differences in gender and age covariate data between the case group and the control group, using propensity score matching to reduce the selection bias between the two groups, and using a multimodal data algorithm to establish a risk prediction assessment model. All patients in the case and control groups were matched 1:1 by age and gender.
7. The method for constructing a depression risk prediction model according to claim 1, characterized in that: In step S53, the core formula of the residual network is: ; Where x and y are the input and output of the network respectively, represents the learned residual mapping, Represents the set of weight parameters of the model.
8. The method for constructing a depression risk prediction model according to claim 1, characterized in that: In step S53, the core formula of the adversarial network is: ; Among them, G represents the generator, D represents the discriminator, minG represents that the generator G tries to minimize the objective function, maxD represents that the discriminator D tries to maximize the objective function, V(D,G) is the value function of GAN, which represents the adversarial goal between the generator and the discriminator; E represents the expected value operator, Indicates that sample x comes from the real data distribution Pdata(x), log D(x) represents the output of the discriminator D for the real sample x; Indicates that the input factor z comes from the factor distribution , represents the fake data generated by the generator G according to the factor z, Denotes the discriminator D to generate data Output.
