A method, device and system for predicting the risk of Alzheimer's disease

Abstract

The application belongs to the technical field of risk prediction, and particularly relates to an Alzheimer's disease risk prediction method, device and system; the method comprises the following steps: acquiring detection data of a to-be-detected object; inputting baseline data and a time accumulation operator into a first classification model to output a first risk prediction result of the to-be-detected object with the time accumulation operator; inputting follow-up data and the time accumulation operator into a second classification model to output a second risk prediction result of the to-be-detected object with the time accumulation operator; and inputting the first risk prediction result and the second risk prediction result into a full connection layer to output a third risk prediction result of the to-be-detected object in a current period and a future period; the application has high prediction accuracy and high practicability, and through early diagnosis of Alzheimer's disease patients, the application helps to make good preparation and prevention measures in time to slow down disease deterioration, improves the survival probability of the to-be-detected object, and has a good application prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of risk prediction, and particularly relates to an Alzheimer's disease risk prediction method, device and system. BACKGROUND

[0002] In recent years, more and more researchers use artificial intelligence methods to conduct prediction research on Alzheimer's disease (AD). Deep learning (DL) shows great potential in the clinical decision support of AD. The main advantage of DL compared with other shallow learning models is that they can learn the most predictive features directly from the original data of a given labeled example dataset. However, studies have shown that the weights of deep models affect the results through multiple layers of combination, and therefore do not produce clinically meaningful solutions. In order to make the research more interpretable and clinically meaningful, machine learning methods are more advantageous. In addition, in the field of artificial intelligence, using some machine learning methods such as PCA dimensionality reduction and deep learning methods for feature processing will cause the data to lose its original structure, which is not clinically meaningful and interpretable in the medical industry. In the AD prediction scenario, researchers often use baseline data and combine traditional machine learning methods for research, but these methods cannot consider time series data and cannot give stage-by-stage prediction results.

[0003] At present, the current AD classification prediction research focuses on mining features that affect the prediction results. For example, Chinese patent CN116994100A discloses a model training method, device, electronic equipment and storage medium. The method comprises: obtaining a sample fundus image of a user that meets the model training condition, the sample fundus image being labeled with a true disease class of Alzheimer's disease; inputting the sample fundus image into a disease prediction model to be trained; calling the feature screening and blood vessel segmentation layer to perform feature screening and blood vessel segmentation labeling processing on the sample fundus image, to obtain an attention feature map and a blood vessel labeling feature map; calling the feature fusion layer to perform feature fusion processing on the attention feature map and the blood vessel labeling feature map, to obtain a predicted disease probability corresponding to the sample fundus image; based on the true disease class and the predicted disease probability, calculating a loss value of the disease prediction model to be trained; in a case where the loss value is within a preset range, obtaining a disease prediction model for predicting Alzheimer's disease. This patent can improve the prediction accuracy of Alzheimer's disease. However, this method ignores the research on the prediction model itself. For the prediction model itself, the existing technology has the following problems: the current use of longitudinal multi-task learning to study AD risk prediction methods only considers the time correlation of longitudinal multi-task learning, and does not consider the time accumulation effect specific to chronic diseases. SUMMARY

[0004] In view of the deficiencies in the prior art, the present application provides an Alzheimer's disease risk prediction method, device and system, which integrates time accumulation effect into model calculation by designing a time accumulation operator, thereby obtaining more accurate early diagnosis results.

[0005] To achieve the above-mentioned purpose, in the first aspect of the present application, the present application provides an Alzheimer's disease risk prediction method, comprising:

[0006] Obtaining various detection data of the to-be-detected object; the various detection data includes follow-up data and baseline data; the follow-up data includes data collected at multiple follow-up times of the to-be-detected object, and the baseline data includes data collected once when the to-be-detected object enters the queue and no longer collected in the follow-up process thereafter;

[0007] Inputting the baseline data and the time accumulation operator into a first classification model to output a first risk prediction result of the to-be-detected object with the time accumulation operator;

[0008] Inputting the follow-up data and the time accumulation operator into a second classification model to output a second risk prediction result of the to-be-detected object with the time accumulation operator;

[0009] Inputting the first risk prediction result and the second risk prediction result into a full connection layer to output a third risk prediction result of the to-be-detected object in the current and a future period of time.

[0010] In the second aspect of the present application, the present application further provides an Alzheimer's disease risk prediction device, comprising:

[0011] A data acquisition module is configured to obtain various detection data of the to-be-detected object; the various detection data includes follow-up data and baseline data; the follow-up data includes data collected at multiple follow-up times of the to-be-detected object, and the baseline data includes data collected once when the to-be-detected object enters the queue and no longer collected in the follow-up process thereafter;

[0012] A first classification module is configured to input the baseline data and the time accumulation operator into a first classification model to output a first risk prediction result of the to-be-detected object with the time accumulation operator;

[0013] A second classification module is configured to input the follow-up data and the time accumulation operator into a second classification model to output a second risk prediction result of the to-be-detected object with the time accumulation operator;

[0014] A third classification module is configured to input the first risk prediction result and the second risk prediction result into a full connection layer to output a third risk prediction result of the to-be-detected object in the current and a future period of time.

[0015] In a third aspect of the present application, the present application also provides a risk prediction system for Alzheimer's disease, comprising a processor, a communication interface and a memory, which are connected with each other, wherein the memory is used for storing a computer program, the computer program comprises program instructions, and the processor is configured to invoke the program instructions to execute the method according to the first aspect of the present application.

[0016] The present application has the advantages that: the present application designs a time accumulation operator by analyzing the concept characteristics of time accumulation effect, which can integrate the medical knowledge of time accumulation effect into model calculation and model prediction; the present application has high prediction accuracy and high practicability, and through early diagnosis of Alzheimer's disease patients, it is helpful to make good preparation and preventive measures in time to slow down the disease deterioration, improve the survival probability of the detected object, and has good application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A risk prediction method for Alzheimer's disease provided by the embodiment of the present application is shown in the scene diagram;

[0018] Figure 2 A flowchart of the risk prediction method for Alzheimer's disease provided by the embodiment of the present application is shown in the scene diagram;

[0019] Figure 3 An input and output diagram of the second classification model in the embodiment of the present application is shown in the scene diagram;

[0020] Figure 4 An input and output diagram of the risk prediction model in the embodiment of the present application is shown in the scene diagram;

[0021] Figure 5 A structure diagram of the risk prediction device for Alzheimer's disease provided by the embodiment of the present application is shown in the scene diagram;

[0022] Figure 6 A structure diagram of the risk prediction system for Alzheimer's disease provided by the embodiment of the present application is shown in the scene diagram. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] The embodiment of the present application obtains various inspection data of the to-be-predicted object, divides the data according to data sources and data forms, and inputs the data into the first classification model and the second classification model which are constructed; then the disease diagnosis result of the to-be-predicted patient in a future period of time is obtained through the full connection layer.

[0025] Please refer to Figure 1 , Figure 1 The scene schematic diagram of the risk prediction method of Alzheimer's disease provided by an embodiment of the present application is shown. The server can obtain various detection data of the to-be-detected object; input the baseline data and the time accumulation operator into the first classification model, and output the first risk prediction result of the to-be-detected object with the time accumulation operator; input the follow-up data and the time accumulation operator into the second classification model, and output the second risk prediction result of the to-be-detected object with the time accumulation operator; input the first risk prediction result and the second risk prediction result into the full connection layer, and output the third risk prediction result of the to-be-detected object in the current and a future period of time. Figure 2 The flowchart of the risk prediction method of Alzheimer's disease in the embodiment of the present application is shown. As shown in Figure 1 The method comprises the following steps:

[0026] 101, obtaining various detection data of the to-be-detected object; the various detection data comprises follow-up data and baseline data; the follow-up data comprises data collected at multiple follow-up times of the to-be-detected object, and the baseline data comprises data collected once when the to-be-detected object enters a queue and no longer collected in the follow-up process thereafter;

[0027] In the embodiment of the present application, various detection data of multiple to-be-detected objects can be obtained. In order to ensure the reliability of the various detection data, the channels for obtaining the various detection data can include but are not limited to public websites, medical institution databases and the like. The various detection data can be obtained through three types of data, i.e., demographic baseline data, imaging data and laboratory examination results. The three types of data are obtained according to different fields and document contents (such as general information, chief complaint, operation record, daily disease course) containing various target (for example, diagnosis, operation, symptom, etc.) models.

[0028] In some possible embodiments, a dataset that conforms to the actual production environment can be selected to complete the AD risk prediction. In the embodiment of the present application, the ADNI database is selected. The Alzheimer's Disease Neuroimaging Initiative (ADNI) is a longitudinal multi-center study aimed at developing clinical, imaging, genetic and biochemical biomarkers for early detection and tracking of Alzheimer's disease (AD). In some possible embodiments, the data used by the present application includes MRI, scale examination data, plasma data, physical examination data, vital sign data, baseline symptom data. Among them, the MRI data, scale examination data, plasma data and part of the physical examination data are follow-up data, and the rest are baseline data. The MRI data uses the ADNI1: Complete 1Yr3T dataset in the database. The dataset is sampled once every six months, and each patient has three sampling samples. The first examination time of the patient is Time1, the reexamination time after six months is Time2, and the reexamination time after one year is Time3. The rest of the data is selected according to the patients in the MRI dataset. A total of 114 patients are selected.

[0029] In the preferred embodiment of the present application, based on the obtained detection data, the detection data can also be preprocessed; can be extracted by artificial labeling, natural language processing (NLP) and the like to obtain structured data (or formatted data). Data cleaning processing can also be performed to obtain cleaned data. Among them, the data cleaning processing includes any combination of secondary feature processing, screening, denoising, interpolation / interpolation, outlier analysis, clustering and dimensionality reduction.

[0030] In some embodiments, the present application uses FreeSurfer image analysis software to perform MRI cortical reconstruction and volume segmentation to obtain 40 ROI regions and 62 brain regions. In the embodiment of the present application, the total volume of ROI, slice mean value and slice standard deviation are selected as the features of the ROI region, and the number of vertices, surface area, gray matter volume, average thickness, thickness standard deviation, average curvature, Gaussian curvature, folding index and intrinsic curvature are selected as the structural features of the brain region. Finally, a total of 678 MRI features are obtained. After integrating all numerical features, missing value processing and normalization are performed. There are three cases of data missing, missing of samples in follow-up data, continuous value missing and discrete value missing. Among the 114 patients, 57 have sample missing, of which 20 are missing in the second sampling and 37 are missing in the third sampling. For sample missing, the embodiment of the present application uses grouped average value for filling. For the missing of continuous feature values, the grouped average value is also used for filling. For the missing of discrete feature values, the majority value is used for filling. For all continuous feature values, the mean-standard deviation method is used for normalization. Among them, there are 703 follow-up data features and 72 baseline data features.

[0031] In the embodiments of the present application, in order to better obtain the time correlation of multi-task and the time accumulation effect specific to chronic diseases, the embodiments of the present application construct a time accumulation operator according to the follow-up time of follow-up data; the present application considers that the embodiment of the time accumulation effect on the classification prediction model should be that the parameters of each follow-up time period consider the influence of the parameters of the historical follow-up time period; and the influence of the risk prediction results of each historical follow-up time period corresponding to the current follow-up time decreases with the time distance from the current follow-up time, and the sum of the influences of all follow-up times is 1.

[0032] The embodiments of the present application design a time accumulation operator: according to the influence size of the detection result of each follow-up time on the detection result of another follow-up time, the weight of each follow-up time relative to another follow-up time is calculated; according to the weight of each follow-up time relative to another follow-up time, the time accumulation operator under each follow-up time is constructed. It can be specifically expressed as:

[0033]

[0034] s.t.ω k. =1,ω∈[0,1]

[0035] Wherein, T represents the time accumulation operator, ω 1,2 represents the weight of the first follow-up time relative to the second follow-up time, and the weight reflects the influence of the first follow-up time on the risk prediction result of the second follow-up time.

[0036] In some embodiments of the present application, the follow-up time of the data used by the present application is 36 months, and sampling is performed once every 12 months, a total of 3 times. The design of the time accumulation operator mainly considers three aspects: current time weight, time window size and distance attenuation:

[0037] Design=(a,d,f)

[0038] Wherein, a represents the size of the current time weight, that is, the influence size of the time accumulation effect on the prediction result; d represents the size of the time window, the larger the time window, the larger the influence range of the historical time considered by the time accumulation effect on the current time; f represents whether the weight attenuates, that is, whether the weight in the time window will attenuate with the increase of the distance from the current time. The time accumulation operator used in the embodiments of the present application is Design=(0.9,3,false), and the formula is:

[0039]

[0040] Based on the time accumulation operator of the above embodiment, the present invention can fully consider the cumulative effect of follow-up data at different time points and improve the accuracy of longitudinal prediction. By weighted integration of the test results of historical time periods, the time accumulation operator can capture the dynamic changes in the patient's disease course and provide more comprehensive information support for disease risk prediction in the current and future period. In addition, by flexibly setting the time window size and weight decay function, the operator ensures that the model can still effectively utilize early follow-up data within a longer follow-up period, thereby achieving more accurate risk prediction.

[0041] 102. Input the baseline data and the time accumulation operator into a first classification model, and output a first risk prediction result of the subject to be detected with the time accumulation operator;

[0042] In an embodiment of the present invention, the first classification model is a decision tree baseline data auxiliary model; the baseline data is input into the first classification model to obtain an initial risk detection result; the initial risk detection result is copied to obtain a zeroth risk detection result with the same dimension as the time accumulation operator; the zeroth risk detection result is combined with the time accumulation operator to obtain a first risk detection result with a time accumulation operator; the baseline data includes the number of samples and the feature dimension; the first risk detection result includes the number of samples, the follow-up time and the number of types.

[0043] Specifically, the present invention takes into account the output of baseline data Since the baseline data has no time dimension, the input and output forms of the baseline data are a two-dimensional tensor, that is, the number of samples and the feature dimension. The input data X and the model parameter matrix W are input into the first classification model to obtain the corresponding output data Y b ∈R n×c , where the input data X is the baseline data, and the model parameter matrix W is the parameter matrix of the first classification model; in order to fuse the output of the first classification model with the output of the second classification model, it is necessary to combine the output with the time accumulation operator to obtain the same as Y f The same form of baseline data auxiliary model output Y b ∈R n×t×c The decision tree model used in the present invention is implemented based on sklearn. The selected decision tree model uses mean square error as the feature selection criterion, randomly selects feature partition points, and has a maximum tree depth of 3.

[0044] like Figure 3 As shown, Figure 3is the input-output structure schematic diagram of the first classification model of the embodiment of the application, it can be found that the initial risk detection result (c1, c2, c3) is copied, the number of copying is the number n of follow-up times, so that a tensor matching the time dimension can be generated, in the tensor, the value of each feature dimension on different follow-up times t1, t2, t3 remains unchanged; that is, the zeroth risk detection result matching the time accumulation operator dimension can be obtained; the zeroth risk detection result is combined with the time accumulation operator T to obtain the first risk detection result with the time accumulation operator; the time accumulation operator captures the cumulative effect of different follow-up times in the time dimension, and after matrix operation, the obtained first risk detection result contains the cumulative effect of each follow-up time, so that the output risk prediction result can comprehensively consider the cumulative effect of baseline data and follow-up data, and provide more accurate risk prediction result and risk prediction basis.103、the follow-up data and the time accumulation operator are input into the second classification model, and the second risk prediction result of the to-be-detected object with the time accumulation operator is output;

[0045] In the embodiment of the application, the follow-up data is respectively matrix cross-multiplied with each classification parameter matrix in the model parameter matrix to obtain a multi-classification matrix of each follow-up data; the multi-classification matrix of each follow-up data is point-multiplied with the time accumulation operator and horizontally summed to obtain a multi-classification time accumulation matrix of each follow-up data; the multi-classification time accumulation matrix of each follow-up data is spliced to obtain a classification prediction result of each follow-up data; the classification prediction results of each follow-up data are integrated to obtain the second risk prediction result of the to-be-detected object with the time accumulation operator; the follow-up data includes the sample number, the feature dimension and the follow-up time; and the second risk prediction result includes the sample number, the follow-up time and the type number.

[0046] Specifically, the second classification model of the embodiment of the application is a longitudinal multi-task follow-up data model; the longitudinal multi-task learning used regards the prediction task of each time point of the sample as a risk task, and requires that the input data form is a three-dimensional tensor, that is, in addition to the sample number and the feature dimension, a one-dimensional follow-up number is added, that is, The model parameter matrix is also a three-dimensional tensor, that is, the classification number, the feature dimension and the follow-up number, the input data X, the model parameter matrix W and the time accumulation operator T are input into the second classification model to obtain the corresponding output data Y f Here, the input data X is the follow-up data including the sample number, the feature dimension and the follow-up time, and the model parameter matrix W is the parameter matrix of the second classification model.

[0047] Specifically, the calculation process includes that the input data is respectively matrix cross-multiplied with each classification parameter matrix Matrix cross product to get the c-classification matrix of n samples Matrix Matrix dot product with time accumulation operator T and transverse summation to get the c-classification time accumulation output of n samples Time accumulation output of c-classification of each sample Concatenation to get the classification prediction matrix of each sample The output of the entire longitudinal multi-task follow-up data model is

[0048] It can be understood that the classification model learns through the training set to establish a mapping from the input space X to the output space Y (discrete value). The principle is to learn a classification function or classification model based on the existing training data set, which can map new input data to predefined categories. The first classification model and the second classification model of the embodiment of the application can be existing classification models, such as logistic regression, decision tree, support vector machine (SVM), random forest, neural network, etc. The application mainly adjusts the multi-input and multi-output tensor of the second classification model, integrates the time accumulation operator, and learns the cumulative effect of each detection data in time, so as to obtain more accurate early diagnosis results.

[0049] 104, input the first risk prediction result and the second risk prediction result into the full connection layer, and output the third risk prediction result of the to-be-detected object in the current and future period of time.

[0050] In the embodiment of the application, the output of the longitudinal multi-task follow-up data model and the decision tree baseline data auxiliary model is fused by using the full connection layer.

[0051] In the embodiment of the application, the risk prediction results are sequentially connected according to the follow-up time points; the first risk prediction result and the second risk prediction result of the current follow-up time and the first risk prediction result and the second risk prediction result of the previous follow-up time are integrated by weighting through a preset weight matrix. The preset weight matrix can be obtained by a trained perception machine; the perception machine can be obtained by training two two-dimensional matrices, i.e. the first risk prediction result and the second risk prediction result.

[0052] The output of the longitudinal multi-task follow-up data model and the decision tree baseline data auxiliary model is combined as the full connection layer input X of the application by using the full connection layer fc = (Y f , Y bThe process of the full connection layer includes: considering that the weights between different models at different times can not be the same, the full connection layer does not directly perform full connection weighted summation on the outputs of the two models, but performs full connection by time, that is, the full connection layer parameters are The training of the full connection layer uses mean square error as a loss function, and uses a stochastic gradient descent algorithm for optimization, and finally obtains the output Y of the AD risk prediction model with time accumulation effect, that is, the third risk prediction result.

[0053] In the preferred embodiment of the present application, considering that medical structures can produce different types and different standards of AD detection data, most of the current AD risk prediction only uses one of the follow-up data or baseline data for classification prediction; and each detection data has different sources and different structures, in order to process these heterogeneous detection data, Figure 4 is the input and output diagram of the risk detection model of the embodiment of the present application, as shown in Figure 4 Each training module can correspond to a risk prediction model. The risk prediction model includes a first classification model, a second classification model and a full connection layer; the model is Figure 4 The overall structure of the risk detection model of the present application is shown in The model first inputs different types of detection data (such as MRI data, scale data, and physiological data) into multiple independent training modules (Trainer 1, Trainer 2, Trainer n) respectively. Each training module is independently trained according to the input detection data type, and outputs the corresponding prediction result. In this process, the time accumulation operator T is also introduced into each training module to enhance the model's processing ability for time dimension information. Then, these prediction results are integrated through a decision convergence strategy. The fusion strategy will weight different prediction results according to their weights w1, w2,... Wn, and finally output the comprehensive prediction result Y finaly This process enables the model to obtain information from different data sources and more accurately predict the risk of Alzheimer's disease through time accumulation effect.

[0054] It can be understood that the preferred embodiment of the present application considers the complexity of data in the actual production environment on the basis of combining the time accumulation operator, trains the multi-source heterogeneous data according to the data source and data form, and provides decision support for doctors. The embodiment solves the problem of single model and each detection data adaptability by dividing and learning the actual inspection data of the to-be-detected object according to the data source and data form; through the improvement of the longitudinal multi-task learning model, the prediction result has the effect of early diagnosis.

[0055] In some embodiments of the present application, the first classification model, the second classification model and the full connection layer constitute a risk prediction model; the model parameters of the first classification model, the second classification model and the full connection layer can be respectively updated according to a loss function, so that the model can predict more accurate risk results. In the embodiments of the present application, the first classification model is optimized by sklearn; the second classification model is optimized by mean square error loss and regularization; and the full connection layer is optimized by mean square error loss.

[0056] The optimization loss function of the present application is constructed as follows:

[0057]

[0058] Wherein, L represents the loss function of the risk prediction model, represents the mean square error loss of the risk prediction model, λ1, λ2 and λ3 are regularization parameters, γ is the risk prediction model, X is the input data matrix, W is the model parameter matrix, T is the time accumulation operator; Y is the output data matrix, ||W||1 is the Lasso coefficient regularization term, is the group sparse regularization term, w kij represents the weight of the i-th feature dimension in the k-th risk task at the j-th follow-up time, c represents the number of risk tasks, i.e. the number of risk types, m represents the feature dimension, t represents the number of follow-up times, and the model optimization process allows joint feature selection and task-specific feature selection for multiple tasks, ||RW T ||1 is the time smoothing regularization term, which considers the time correlation between tasks, and R is the time smoothing matrix.

[0059] The specific values are as follows:

[0060] If the to-be-selected follow-up time i is consistent with the current follow-up time j, the time smoothing matrix Rij takes the value of 1, indicating that the influence of the to-be-selected follow-up time i on the current follow-up time j is self-correlation.

[0061] If the to-be-selected follow-up time i is the previous time point of the current follow-up time j, the time smoothing matrix Rij takes the value of -1, indicating that the influence of the to-be-selected follow-up time i on the current follow-up time j is mutual correlation.

[0062] It can be represented as:

[0063]

[0064] where i represents the follow-up time to be selected. This is the row index of the matrix, representing the features or outputs of the model at a certain time point. j represents the current follow-up time j or the historical follow-up time. This is the column index of the matrix, representing the influence of other time points associated with a certain time point. By adjusting the difference between the current time point and the previous time point, smoothing in the time dimension is achieved. It ensures that the model has a smooth transition over time, avoiding sharp fluctuations between time points, thereby improving the stability of the model and the accuracy of the prediction.

[0065] It can be understood that the present application controls the weight of the current time point when self-associating, and the cross-association is used to adjust the influence of the previous time point on the current time point, ensuring a smooth transition in the time dimension.

[0066] In some embodiments, the first classification model can be optimized directly through sklearn. The second classification model and the fully connected layer can also include other losses, such as mean absolute error, cross-entropy loss, etc.

[0067] Based on the above process, the present application divides each detection data of the to-be-detected object, inputs it into the risk prediction model, and obtains the disease prediction output through the first classification model, the second classification model, and the fully connected layer where t represents the predictable future time, and c represents the disease development, which has three types for Alzheimer's disease: normal cognitive level (CN), mild cognitive impairment (MCI), and Alzheimer's disease (AD).

[0068] In some embodiments of the present application, the present application also provides a risk prediction device for Alzheimer's disease, as shown in Figure 5 The risk prediction device 200 includes:

[0069] The data acquisition module 201 is configured to acquire each detection data of the to-be-detected object; the detection data includes follow-up data and baseline data; the follow-up data includes data collected at multiple follow-up times of the to-be-detected object, and the baseline data includes data collected once when the to-be-detected object enters the queue and no longer collected in the subsequent follow-up process;

[0070] The first classification module 202 is configured to input the baseline data and the time accumulation operator into the first classification model, and output the first risk prediction result of the to-be-detected object with the time accumulation operator;

[0071] The second classification module 203 is configured to input the follow-up data and the time accumulation operator into the second classification model, and output the second risk prediction result of the to-be-detected object with the time accumulation operator;

[0072] The third classification module 204 is configured to input the first risk prediction result and the second risk prediction result into a full connection layer, and output a third risk prediction result of the to-be-detected object in the current time and a future period of time.

[0073] The risk prediction device for Alzheimer's disease can be integrated in a computer device, which can be a terminal or a server, etc. The terminal can be a smart phone, a tablet computer, a notebook computer, a personal computer (PC) or the like. The server can be a physical server, a server cluster composed of multiple physical servers or a distributed system, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, etc. The terminal and the server can be connected directly or indirectly through wired or wireless communication, which is not limited in the present application. In some embodiments of the present application, a risk prediction system 300 for Alzheimer's disease is also provided, as shown in Figure 6 The risk prediction system 300 for Alzheimer's disease can include at least one processor 301, at least one memory 302 and a corresponding communication interface 303. Those skilled in the art can understand that the risk prediction system 300 for Alzheimer's disease shown in Figure 6 does not constitute a limitation on the system, and can include more or fewer components than those shown, or combine some components, or different component arrangements. Among them:

[0074] The processor 301 is the control center of the risk prediction system 300 for Alzheimer's disease, which executes various functions of the risk prediction system 300 for Alzheimer's disease and processes data by running or executing software programs and / or modules stored in the memory 302 and calling data stored in the memory 302, thereby monitoring the risk prediction system 300 for Alzheimer's disease as a whole. Optionally, the processor 301 can include one or more processing cores; preferably, the processor 302 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface and application programs, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 301.

[0075] The data of each patient in the follow-up data set corresponds to a matrix with a size of 3*703, 3 rows represent sampling at 3 time points, and 703 represents 703 preprocessed features; the optimization algorithm selected for the longitudinal multi-task follow-up data model is the Nesterov accelerated gradient descent method, the learning rate is 0.01, the weight decay is 0.1, and the momentum is 0.1; the decision tree baseline data assisted model is realized based on sklearn, and the selected decision tree model uses mean square error as the feature selection standard, randomly selects the feature division point, and the maximum tree depth is 3. The entire training iteration round is 5000 rounds.

[0076] The following are specific experimental results, comparing the evaluation parameters of the present application and other traditional machine learning algorithms and their algorithm combinations. As shown in Table 1:

[0077] Table 1 Comparison of accuracy of six time accumulation operators designed by the present application on the longitudinal multi-task model

[0078]

[0079]

[0080] For the ADNI data set, the average accuracy of the present application on each time group reached 76.83%, which was the best among all model combinations, followed by the MTL+KNN model combination. In a real medical scenario, different data set distributions and feature attributes differ, resulting in different optimal model combinations. And from the experimental results, it can be analyzed that the average accuracy of the algorithm model using the time accumulation operator on each time group was improved by 22.78%, and the KNN algorithm and the LR algorithm obtained a larger improvement, especially in the prediction of the Time3 group, the accuracy was improved most obviously, proving the effectiveness of the time accumulation operator and the superiority of the present application in early diagnosis and prediction of Alzheimer's disease. The present application is implemented on the ADNI data set, proving that the present application is superior to traditional machine learning single models and other models that do not consider the time accumulation effect in terms of the correctness of early diagnosis and prediction results of AD, model generalization ability, and the like.

[0081] It should be noted that the present application proposes an Alzheimer's disease risk prediction method, device and system to address the shortcomings of existing early diagnosis and prediction methods for Alzheimer's disease, applies the time accumulation effect to the AD risk prediction scheme, designs a time accumulation operator from three angles by analyzing the conceptual characteristics of the time accumulation effect, and integrates it into the risk prediction model calculation, and the experimental results prove the effectiveness and superiority of the time accumulation operator designed by the present application.

[0082] The above examples further illustrate the objects, technical solutions and advantages of the present application. It should be understood that the above examples are only preferred embodiments of the present application and are not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made to the present application within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting the risk of Alzheimer's disease, characterized by, The method comprises the following steps: acquiring detection data of a to-be-detected object; the detection data comprises follow-up data and baseline data; the follow-up data comprises data collected at multiple follow-up times, and the baseline data comprises data collected once when the to-be-detected object enters a queue and no longer collected in the follow-up process; inputting the baseline data and a time accumulation operator into a first classification model to output a first risk prediction result of the to-be-detected object with the time accumulation operator; the calculation process of the time accumulation operator comprises calculating a weight of each follow-up time relative to another follow-up time according to an influence size of a detection result of each follow-up time on a detection result of another follow-up time; constructing a time accumulation operator at each follow-up time according to the weight of each follow-up time relative to another follow-up time; inputting the baseline data and the time accumulation operator into the first classification model to output the first risk prediction result of the to-be-detected object with the time accumulation operator comprises inputting the baseline data into the first classification model to obtain an initial risk detection result; copying the initial risk detection result to obtain a zeroth risk detection result with the same dimension as the time accumulation operator; combining the zeroth risk detection result and the time accumulation operator to obtain the first risk detection result with the time accumulation operator; the baseline data comprises a sample number and a feature dimension; the first risk detection result comprises a sample number, a follow-up time and a type number; inputting the follow-up data and the time accumulation operator into a second classification model to output a second risk prediction result of the to-be-detected object with the time accumulation operator; inputting the follow-up data and the time accumulation operator into the second classification model to output the second risk prediction result of the to-be-detected object with the time accumulation operator comprises: performing matrix cross multiplication of the follow-up data and each classification parameter matrix in a model parameter matrix to obtain a multi-classification matrix of each follow-up data; performing point multiplication of the multi-classification matrix of each follow-up data and the time accumulation operator and performing horizontal summation to obtain a multi-classification time accumulation matrix of each follow-up data; splicing the multi-classification time accumulation matrix of each follow-up data to obtain a classification prediction result of each follow-up data; integrating the classification prediction result of each follow-up data to obtain the second risk prediction result of the to-be-detected object with the time accumulation operator; the follow-up data comprises a sample number, a feature dimension and a follow-up time; and the second risk prediction result comprises a sample number, a follow-up time and a type number. inputting the first risk prediction result and the second risk prediction result into a full connection layer to output a third risk prediction result of the to-be-detected object in the current and a future period of time; the inputting the first risk prediction result and the second risk prediction result into the full connection layer to output the third risk prediction result of the to-be-detected object in the current and the future period of time comprises performing a step-by-step full connection operation on the risk prediction results according to follow-up times; and the first risk prediction result and the second risk prediction result of the current follow-up time and the first risk prediction result and the second risk prediction result of a plurality of previous follow-up times are weighted and integrated through a preset weight matrix. The first classification model, the second classification model and the full connection layer constitute a risk prediction model; and the risk prediction model is based on mean square error loss optimization and regularization optimization.

2. The method of claim 1, wherein the risk of Alzheimer's disease is predicted by the step of: The risk prediction model based on mean square error loss optimization and regularization optimization is represented as: wherein, is a loss function of the risk prediction model, is a mean squared error loss of the risk prediction model, , and is a regularization parameter, is a risk prediction model, is an input data matrix, is a model parameter matrix, is a time accumulation operator; is an output data matrix, is a Lasso coefficient regularization term, is a group sparsity regularization term, represents the weight of the i-th feature dimension in the k-th risk task at the j-th follow-up time, represents the number of risk tasks, i.e., the number of risk types, represents the feature dimension, represents the number of follow-up times, is a time-smoothing regularization term, is a time-smoothing matrix.

3. The method of claim 2, wherein the risk of Alzheimer's disease is predicted by the step of: The time smoothing matrix has the following specific values: when the to-be-selected follow-up time is consistent with the current follow-up time, the time smoothing matrix takes a value of 1, indicating that the influence of the to-be-selected follow-up time on the current follow-up time is self-association; and when the to-be-selected follow-up time is a time point before the current follow-up time, the time smoothing matrix takes a value of -1, indicating that the influence of the to-be-selected follow-up time on the current follow-up time is mutual association. The device comprises:

4. An apparatus for predicting the risk of Alzheimer's disease, which is used to execute the method for predicting the risk of Alzheimer's disease according to any one of claims 1 to 3, characterized by a data acquisition module configured to acquire various detection data of a to-be-detected object; the various detection data comprises follow-up data and baseline data; the follow-up data comprises data collected at a plurality of follow-up times of the to-be-detected object, and the baseline data comprises data collected once when the to-be-detected object enters a queue and no longer collected in a follow-up process thereafter; a first classification module configured to input the baseline data and a time accumulation operator into a first classification model to output a first risk prediction result of the to-be-detected object with the time accumulation operator; a second classification module configured to input the follow-up data and the time accumulation operator into a second classification model to output a second risk prediction result of the to-be-detected object with the time accumulation operator; a third classification module configured to input the first risk prediction result and the second risk prediction result into a full connection layer to output a third risk prediction result of the to-be-detected object in the current and a future period of time. The device comprises a processor, a communication interface and a memory, which are connected to each other, wherein the memory is configured to store a computer program, the computer program comprises program instructions, and the processor is configured to invoke the program instructions to execute the method according to any one of claims 1 to 3.

5. A system for predicting the risk of Alzheimer's disease, characterized by, ​

Images