A multi-dimensional colorectal cancer risk prediction method, medium and device

Abstract

The present invention relates to a colorectal cancer risk prediction method based on multi-dimensionality. The LSTM network model is used to predict the user's life status changes in the future according to the user's current actual life information. In addition, the risk prediction method uses a DBN probability algorithm model to integrate the colorectal disease information of the user's family to generate the prediction characteristics of the user's own colorectal disease probability distribution. The colorectal cancer risk prediction method based on multi-dimensionality realizes the dynamic prediction of the user's future colorectal cancer risk by combining the user's actual life information and family disease information. This dynamic, multi-dimensional risk prediction result is more accurate and comprehensive than the existing method, realizes the dynamic and accurate prediction of colorectal cancer risk, provides users with more targeted health management suggestions, and has strong practical value.

Description

Technical Field

[0001] The present application relates to the field of smart medicine, and specifically to a multi-dimensional colorectal cancer risk prediction method, medium and device. Background Art

[0002] Colorectal cancer is a malignant tumor that occurs in the colon and rectum. It usually originates from the abnormal proliferation and differentiation of the intestinal mucosal epithelial cells and gradually develops into a tumor. Colorectal cancer includes colon cancer and rectal cancer. Colon cancer mainly occurs in different parts of the colon, such as the ascending colon, transverse colon, descending colon and sigmoid colon, and rectal cancer occurs in the rectum segment between the anus and the sigmoid colon. Most colorectal cancers are caused by the adenoma-malignancy sequence of intestinal epithelial cells. Colorectal cancer is also associated with a genetic history. Diagnostic methods include colonoscopy, imaging examinations and clinical symptoms, combined with doctors to conduct specific disease assessment, prediction and diagnosis.

[0003] Colorectal cancer generally evolves from common intestinal diseases, that is, patients with long-term enteritis have a higher probability of developing colorectal cancer. The occurrence of enteritis is mostly affected by the user's usual diet, work and rest, etc. The existing colorectal cancer risk prediction method can analyze the probability of cancer based on the user's living habits, and even introduce genetic factors, combining genetic factors with living habits for further analysis of the probability of cancer. However, the risk of colorectal cancer is dynamically changing, that is, the risk prediction results currently output can only show the cancer risk in the current state, and do not consider the impact of the time dimension, resulting in the existing cancer risk result prediction is not accurate enough. Summary of the invention

[0004] In view of the above problems, the present application provides a multi-dimensional colorectal cancer risk prediction method, medium and device, which solves the problem that the existing colorectal cancer risk prediction is not accurate enough.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for predicting colorectal cancer risk based on multiple dimensions, comprising:

[0006] Acquire the actual life information of the current user in the first preset time period, the actual life information includes actual diet information, actual work and rest information, actual behavior information and actual gastrointestinal information, the actual diet information includes actual diet style, actual diet frequency, actual diet category and actual diet duration, the actual work and rest information includes actual sleep duration, actual sleep time period, actual wake-up time period and actual sleep quality, the actual behavior information includes actual exercise information, actual sitting and standing information and actual walking information, the actual gastrointestinal information includes actual defecation duration, actual defecation status, actual stomachache status, actual stomachache duration, actual enteritis duration and actual enteritis status;

[0007] Representing real life information as a first multi-dimensional vector, wherein the first multi-dimensional vector includes a plurality of first information vectors of different dimensions;

[0008] Input the first multidimensional vector into the LSTM network model to obtain a first prediction feature, which is obtained by fusion of the LSTM network model through the attention mechanism. The first prediction feature is the predicted life information of the current user in the second preset time period, and the predicted life information includes predicted diet information, predicted work and rest information, predicted behavior information, and predicted gastrointestinal information;

[0009] Obtain the family disease information of the current user, which includes the colorectal disease information of multiple family members, including the time of illness, disease category, course of illness, and recovery status;

[0010] Representing the family disease information as a second multi-dimensional vector, wherein the second multi-dimensional vector includes a plurality of second information vectors of different dimensions;

[0011] Inputting the second multidimensional vector into the DBN probability algorithm model to obtain a second prediction feature, where the second prediction feature is probability distribution information of the disease category of the colorectal disease of the current user;

[0012] The first prediction feature and the second prediction feature are input into the trained neural network model to obtain a risk prediction result, which is the risk probability change information of the current user who may suffer from colorectal diseases within the second preset time period and the disease category of the colorectal diseases that may be suffered.

[0013] In some embodiments, the real life information is represented as a first multi-dimensional vector, and the first multi-dimensional vector includes a plurality of first information vectors of different dimensions including:

[0014] The actual life information is grouped in sequence according to the actual diet information, the actual work and rest information, the actual behavior information and the actual gastrointestinal information to obtain the first group, the second group, the third group and the fourth group. The first group is the actual diet information, the second group is the actual work and rest information, the third group is the actual behavior information, and the fourth group is the actual gastrointestinal information.

[0015] Extracting non-numerical information in the first group to obtain first category information, performing one-hot encoding on the first category information to obtain a first encoding vector, and extracting numerical information in the first group to obtain first indicator information, performing a logical operation on the first indicator information to obtain a first feature scalar corresponding to the first encoding vector, and combining the first encoding vector with the first feature scalar to form a first information vector of a first dimension;

[0016] Extracting the non-numerical information in the second group to obtain second category information, performing one-hot encoding on the second category information to obtain a second encoding vector, and extracting the numerical information in the second group to obtain second indicator information, performing a logical operation on the second indicator information to obtain a second feature scalar corresponding to the second encoding vector, and combining the second encoding vector with the second feature scalar to form a first information vector of a second dimension;

[0017] Extracting the non-numerical information in the third group to obtain third category information, performing one-hot encoding on the third category information to obtain a third encoding vector, and extracting the numerical information in the third group to obtain third indicator information, performing a logical operation on the third indicator information to obtain a third feature scalar corresponding to the third encoding vector, and combining the third encoding vector with the third feature scalar to form a first information vector of a third dimension;

[0018] Extracting the non-numerical information in the fourth group to obtain fourth category information, performing one-hot encoding on the fourth category information to obtain a fourth encoding vector, and extracting the numerical information in the fourth group to obtain fourth indicator information, performing a logical operation on the fourth indicator information to obtain a fourth characteristic scalar corresponding to the fourth encoding vector, and combining the fourth encoding vector with the fourth characteristic scalar to form a first information vector of a fourth dimension;

[0019] The first information vector of the first dimension, the first information vector of the second dimension, the first information vector of the third dimension, and the first information vector of the fourth dimension are concatenated to form a first multi-dimensional vector.

[0020] In some embodiments, the first information vector of multiple dimensions includes multiple first feature points, the LSTM network model includes an encoder module and a decoder module, and the first prediction feature includes multiple second feature points;

[0021] The first multidimensional vector is input into the LSTM network model, and the first prediction feature is obtained including:

[0022] Arrange multiple first feature points in the same first information vector according to a time series, and segment the multiple first feature points according to a preset time step to obtain a first data set;

[0023] Inputting the first feature points in the first data set into the encoder module one by one according to a preset time step;

[0024] The encoder module performs encoding operation according to the plurality of first feature points to obtain a first hidden state corresponding to each first feature point, and inputs each first hidden state into the attention weight algorithm model to obtain a first attention weight corresponding to each first hidden state;

[0025] Performing weighted summation on the multiple first hidden states according to the first attention weight to obtain a first context vector;

[0026] Inputting the first context vector into a decoder module;

[0027] The decoder module generates a second hidden state of a plurality of second feature points divided according to a preset time step according to the first context vector;

[0028] Input each second hidden state into the attention weight algorithm model to obtain a second attention weight corresponding to each second hidden state;

[0029] Performing weighted summation on the multiple second hidden states according to the second attention weight to obtain a first information feature;

[0030] Inputting the first information feature and the plurality of second hidden states into a fully connected layer to obtain a second data set, wherein the second data set includes a plurality of second feature points;

[0031] Repeat the above steps until each dimension in the first multidimensional vector has a second data set;

[0032] Features of the multiple second data sets are concatenated to form a first prediction feature.

[0033] In some embodiments, the encoder module includes a plurality of first LSTM units, the first LSTM units include a first input gate, a first forget gate, and a first output gate, the decoder module includes a plurality of second LSTM units, the second LSTM units include a second input gate, a second forget gate, and a second output gate;

[0034] The encoder module performs encoding operation according to the first feature point currently input, and obtains the first hidden state corresponding to each first feature point including:

[0035] The following steps are performed for each first feature point input to the encoder module one by one:

[0036] Reading a first feature point corresponding to a current preset time step, concatenating a first initial input value of the first feature point with a first hidden state of the first feature point corresponding to a previous preset time step, to obtain a first final input value of the first feature point corresponding to the current preset time step;

[0037] Calculating an activation value of a first input gate, an activation value of a first forget gate, and an activation value of a first output gate according to the first final input value;

[0038] Update the first cell state of the first feature point corresponding to the current preset time step;

[0039] Calculate a first hidden state of a first feature point corresponding to a current preset time step according to an activation value of a first input gate, an activation value of a first forget gate, an activation value of a first output gate, and a first cell state;

[0040] The decoder module generates a second hidden state of a plurality of second feature points divided according to a preset time step according to the first context vector, including:

[0041] The first context vector is used as the second hidden state of the first second feature point, and the second cell state of the first second feature point is initialized to zero;

[0042] The following steps are performed for each second feature point input to the decoder module one by one:

[0043] Reading an output value of the second feature information corresponding to the previous preset time step as a second initial input value of the second feature point corresponding to the current preset time step, and concatenating the second initial input value with the first context vector to obtain a second final input value of the second feature point corresponding to the current preset time step;

[0044] Calculating an activation value of a second input gate, an activation value of a second forget gate, and an activation value of a second output gate according to the second final input value;

[0045] Updating the second cell state of the second feature point corresponding to the current preset time step;

[0046] The second hidden state of the second feature point corresponding to the current preset time step is calculated according to the activation value of the second input gate, the activation value of the second forget gate, the activation value of the second output gate, and the second cell state.

[0047] In some embodiments, the family disease information is represented as a second multidimensional vector, and the second multidimensional vector includes a plurality of second information vectors of different dimensions including:

[0048] Dividing the plurality of colorectal disease information in the family disease information by family members to obtain disease information of each family member;

[0049] Perform the following steps for each family member's medical condition information:

[0050] Extracting non-numerical information from the family member's symptom information to obtain first symptom information;

[0051] Performing one-hot encoding on the first symptom information to obtain a first symptom vector;

[0052] and extracting the numerical information from the family member's symptom information to obtain second symptom information;

[0053] Performing a logical operation on the second symptom information to obtain a first symptom scalar corresponding to the first symptom vector;

[0054] combining the first symptom vector with the first symptom scalar to form a second information vector;

[0055] The plurality of second information vectors are concatenated to form a second multi-dimensional vector.

[0056] In some embodiments, the DBN probability algorithm model includes multiple RBM layers stacked from top to bottom, each RBM layer includes visible units and hidden units, and the multiple RBM layers are numbered in sequence from top to bottom in the stacking order, the topmost RBM layer is recorded as the first RBM layer, and the bottommost RBM layer is recorded as the Nth RBM layer. The second multidimensional vector is input into the DBN probability algorithm model, and the second prediction feature is obtained, including:

[0057] The second multidimensional vector is used as the input feature of the visible unit of the first RBM layer;

[0058] The first RBM layer generates the feature code of the hidden unit of the first RBM layer according to the second multidimensional vector, which is recorded as the first feature code and expressed by formula (1). Formula (1) is as follows:

[0059] ;

[0060] In formula (1), is the first feature code, is the activation function of the first RBM layer, is the weight matrix of the first RBM layer, is the bias vector of the first RBM layer;

[0061] The first feature code is used as the input feature of the visible unit of the second RBM layer to generate the second feature code of the second RBM layer, which is expressed by formula (2). Formula (2) is as follows:

[0062] ;

[0063] In formula (2), For the The input features of an RBM are Encode the features of the previous RBM layer, , For the The activation function of the RBM layer is For the The weight matrix of each RBM layer is For the The bias vector of each RBM layer;

[0064] Repeat the above steps until the Nth RBM layer is reached, and the Nth feature code of the Nth RBM layer is obtained. The Nth feature code is used as the second prediction feature and substituted into formula (2) The second prediction feature is obtained, which is expressed as formula (3). Formula (3) is as follows:

[0065] ;

[0066] In formula (3), is the second prediction feature, For the The input features of an RBM are: For the The activation function of the RBM layer is For the The weight matrix of each RBM layer is For the The bias vector of the RBM layer.

[0067] In some embodiments, the method further comprises:

[0068] Generate a first disease curve of multiple colorectal diseases within a second preset time period according to the risk prediction result and the first prediction feature, where the first disease curve is a disease probability change curve predicted by the current user;

[0069] and, selecting in the database a plurality of first body vectors corresponding to a first group whose vector similarity is in the same vector similarity range as that of the first multidimensional vector;

[0070] Obtaining a second disease curve corresponding to each first body vector;

[0071] Matching the slope similarities of the first disease curve and the second disease curve one by one, dividing the individuals of the second disease curve in the same slope similarity range to obtain a second group, and recording the first body vector in the second group as the second body vector;

[0072] Obtaining a second intervention curve corresponding to each second individual vector, where the second intervention curve is an actual disease probability change curve of the second group after receiving the corresponding second intervention scheme;

[0073] generating a primary intervention plan corresponding to the current user according to the plurality of second intervention curves;

[0074] The fine-tuning information is obtained, where the fine-tuning information is generated according to the manual input of the user, and the primary intervention plan is adjusted according to the fine-tuning information to obtain the final intervention plan for the current user.

[0075] In some embodiments, generating a primary intervention plan corresponding to the current user according to the plurality of second intervention curves includes:

[0076] Calculating the slope difference between each second intervention curve and its corresponding second disease curve;

[0077] Arrange the multiple slope differences from large to small, and select the second intervention plan corresponding to the second intervention curve with the largest slope difference as the primary intervention plan.

[0078] In a second aspect, the present invention further provides a computer-readable storage medium storing computer program instructions, which implement the method described in the first aspect when executed by a processor.

[0079] In a third aspect, the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method described in the first aspect.

[0080] Different from the prior art, in the above technical solution, the risk prediction method uses the LSTM network model to predict the user's life status changes in the future period of time according to the user's current actual life information (diet, work and rest, behavior, gastrointestinal, etc.), which provides important basic data for the subsequent colorectal cancer risk prediction. In addition, the risk prediction method uses the DBN probability algorithm model to integrate the colorectal disease information of the user's family, and generates the prediction characteristics of the probability distribution of the user's own colorectal disease, which can more comprehensively evaluate the user's genetic risk factors. The above two prediction features are input into the trained neural network model, which can not only obtain the risk probability changes of the user's future colorectal cancer, but also predict the specific type of colorectal disease that may be suffered. This technical solution is based on a multi-dimensional colorectal cancer risk prediction method. By combining the user's actual life information and family disease information, it realizes the dynamic prediction of the user's future colorectal cancer risk. This dynamic, multi-dimensional risk prediction result is more accurate and comprehensive than the existing method, and realizes the dynamic and accurate prediction of colorectal cancer risk, providing users with more targeted health management suggestions, and has a strong practical value.

[0081] The above-mentioned records related to the invention content are only an overview of the technical solution of the present application. In order to enable ordinary technicians in the field to more clearly understand the technical solution of the present application, and then implement it according to the text of the specification and the contents recorded in the drawings, and to make the above-mentioned purpose and other purposes, features and advantages of the present application easier to understand, the following is an explanation in combination with the specific implementation mode and drawings of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The drawings are only used to illustrate the principles, implementation methods, applications, characteristics and effects of the specific embodiments of the present invention and other related contents, and shall not be considered as limiting the present application.

[0083] In the drawings of the specification:

[0084] Figure 1 It is a schematic diagram of steps S101 to S107 of the risk prediction method described in the specific implementation method;

[0085] Figure 2 It is a schematic diagram of steps S201 to S211 of the risk prediction method described in the specific implementation method;

[0086] Figure 3 It is a schematic diagram of steps S301 to S305 of the risk prediction method described in the specific implementation method;

[0087] Figure 4 It is a schematic diagram of steps S401 to S406 of the risk prediction method described in the specific implementation method;

[0088] Figure 5 A schematic diagram of the second intervention curve and the second disease curve in a specific implementation manner;

[0089] Figure 6 It is a schematic diagram of the change of the slope difference described in the specific implementation method. DETAILED DESCRIPTION

[0090] In order to explain in detail the possible application scenarios, technical principles, specific schemes that can be implemented, and the purposes and effects that can be achieved, the following is a detailed description of the specific embodiments listed in conjunction with the accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of the present application, and are therefore only used as examples, and cannot be used to limit the scope of protection of the present application.

[0091] Reference to "embodiment" herein means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present application. The term "embodiment" appearing in various places in the specification does not necessarily refer to the same embodiment, nor does it particularly limit its independence or association with other embodiments. In principle, in the present application, as long as there is no technical contradiction or conflict, the various technical features mentioned in the embodiments can be combined in any way to form a corresponding implementable technical solution.

[0092] Unless otherwise defined, the technical terms used in this document have the same meanings as those generally understood by those skilled in the art to which this application belongs; the use of relevant terms in this document is only for describing specific embodiments and is not intended to limit this application.

[0093] In the description of this application, the term "and / or" is an expression used to describe the logical relationship between objects, indicating that three relationships may exist, for example, A and / or B, which means: A exists, B exists, and A and B exist at the same time. In addition, the character " / " in this article generally indicates that the objects before and after are in an "or" logical relationship.

[0094] In the present application, terms such as “first” and “second” are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship of quantity, priority or sequence between these entities or operations.

[0095] Without further limitations, in this application, the words "include", "comprises", "has" or other similar open-ended expressions used in sentences are intended to cover non-exclusive inclusion. These expressions do not exclude the presence of additional elements in the process, method or product including the elements, so that the process, method or product including a series of elements may include not only those limited elements, but also other elements not explicitly listed, or also include elements inherent to such process, method or product.

[0096] Similar to the understanding in the Examination Guidelines, in this application, expressions such as "greater than", "less than", "exceed" and the like are understood to exclude the number itself; expressions such as "above", "below", "within" and the like are understood to include the number itself. In addition, in the description of the embodiments of this application, "multiple" means more than two (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups", "multiple times", etc., unless otherwise clearly and specifically limited.

[0097] See also Figure 1 In a first aspect, this embodiment provides a method for predicting colorectal cancer risk based on multiple dimensions, comprising:

[0098] S101, obtaining the actual life information of the current user in the first preset time period, the actual life information including actual diet information, actual work and rest information, actual behavior information and actual gastrointestinal information, the actual diet information including actual diet style, actual diet frequency, actual diet category and actual diet duration, the actual work and rest information including actual sleep duration, actual sleep time period, actual wake-up time period and actual sleep quality, the actual behavior information including actual exercise information, actual sitting and standing information and actual walking information, the actual gastrointestinal information including actual defecation duration, actual defecation status, actual stomachache status, actual stomachache duration, actual enteritis duration and actual enteritis status;

[0099] S102, representing real life information as a first multi-dimensional vector, where the first multi-dimensional vector includes a plurality of first information vectors of different dimensions;

[0100] S103, inputting the first multidimensional vector into the LSTM network model to obtain a first prediction feature, the LSTM network model is obtained by fusion through the attention mechanism, the first prediction feature is the predicted life information of the current user in the second preset time period, and the predicted life information includes predicted diet information, predicted work and rest information, predicted behavior information and predicted gastrointestinal information;

[0101] S104, obtaining family disease information of the current user, the family disease information including colorectal disease information of multiple family members, the colorectal disease information including disease time, disease category, disease course and recovery status;

[0102] S105, expressing the family disease information as a second multidimensional vector, wherein the second multidimensional vector includes a plurality of second information vectors of different dimensions;

[0103] S106, inputting the second multidimensional vector into the DBN probability algorithm model to obtain a second prediction feature, where the second prediction feature is probability distribution information of the disease category of the colorectal disease of the current user;

[0104] S107. Input the first prediction feature and the second prediction feature into the trained neural network model to obtain a risk prediction result, which is the risk probability change information of the current user who may suffer from colorectal disease within the second preset time period and the disease category of the colorectal disease that may be suffered.

[0105] In step S101, the first preset time period can be understood as a pre-set data collection cycle, for example, it can be a collection cycle of the past month, the past three months, and the past year. Actual life information includes actual dietary information, actual work and rest information, actual behavior information, and actual gastrointestinal information. Specifically, the actual dietary information can be collected according to actual needs. For example, when the unit is a day, the actual dietary style in the actual dietary information is the actual dietary style of the day, the actual number of meals on the day, the actual dietary category of the day, and the actual dietary duration of the day. The actual work and rest information, the actual behavior information, and the actual gastrointestinal information can refer to the time period segmentation of the actual dietary information for corresponding data collection. In this embodiment, the actual life information covers the fields of diet, work and rest, behavior, and gastrointestinal, and can realize multi-angle information collection of users, so as to facilitate the subsequent use of actual life information to further predict the health status of users. It can be understood that the means for obtaining actual life information can be carried out by developing a dedicated information collection application for corresponding data collection, and this embodiment does not limit this.

[0106] In step S102, the first information vector can be understood as a specific data feature in actual life information. Preferably, the first multidimensional matrix can be used to replace the first multidimensional vector. Each dimension represents a category in actual life information, and each dimension includes multiple specific data features. Each specific data feature is represented as a first information vector. In this way, a more comprehensive feature embodiment of actual life information can be achieved.

[0107] In step S103, the first multidimensional vector is input into the LSTM network model, which is a long short-term memory network model, a deep learning model commonly used to process sequence data, and is composed of multiple repeated LSTM units, each of which is usually composed of an input gate, a forget gate, and an output gate. When each time step t receives a new input, the input gate will evaluate the importance of the current information and decide whether to retain it. If the forget gate is closed, the cell state of the previous time step will be discarded. The function of the output gate is to determine which parts of the cell state need to be propagated to the output state. In all three gates, an S-shaped activation function is used, and the result range is [0, 1], so as to allow the gate to be in a semi-open state. It can be understood that the LSTM network model obtains and learns the feature change law covered by the data information in the past period of time by forward propagation. Further, the LSTM network model can output the prediction information for a period of time in the future through backward propagation, that is, the feature change law corresponding to the data information of the same category in a certain time segment in the future is given. This embodiment uses the LSTM network model and uses the first multidimensional vector as the input feature of the LSTM network model to obtain the first prediction feature, which is the predicted life information of the current user in the second preset time period in the future, so as to facilitate subsequent further processing. In this embodiment, the LSTM network model also integrates the attention mechanism, avoiding the error problem generated by the traditional LSTM network model in the encoding and decoding process, reducing the deviation area, thereby realizing the refined analysis of multiple types of features in the first multidimensional vector, making the result of the first prediction feature more accurate.

[0108] In step S104, the family disease information of the current user is obtained. The family disease information can be associated with the current user's kinship through the user's identity information, and the colorectal disease information of multiple family members of the current user is further obtained. It should be noted that the colorectal disease information can be understood as a variety of diseases in the colorectal field, which may include polyps, inflammation, benign tumors and malignant tumors, that is, colorectal cancer can belong to one of the diseases in the colorectal field. Specifically, the colorectal disease information includes the time of illness, the type of disease, the course of illness and the state of recovery. Among them, when a single family member has multiple experiences of colorectal disease, the family member can be used as a grouping unit, and a single colorectal disease record can be collected as a colorectal disease information. Further, the time of illness can be understood as the time from the onset of symptoms to the time of recovery, the course of illness is the specific content of the entire disease process, and the state of recovery can be understood as the specific information of the current family member after recovery, which can be healthy, well recovered or occasionally painful, etc., and recorded according to the actual situation.

[0109] In step S105, the family disease information is represented as a second multidimensional vector, which may specifically include the following steps: converting the illness time into a time difference, converting different disease categories into numerical form, and representing the course of disease information and the recovery status in the form of text features. The kinship characteristics of each family member and the current user in the family disease information are stored synchronously. It is understandable that, similar to the first multidimensional vector, a second multidimensional matrix can be used to replace the second multidimensional vector, and each dimension represents the colorectal disease information of a family member in a kinship hierarchy, that is, there can be multiple second information vectors in each dimension.

[0110] In step S106, the second multidimensional vector is input into the DBN probability algorithm model to perform deep learning and evaluation on the impact of family members with different kinship relationships suffering from colorectal diseases on the probability of the current user suffering from the disease. The DBN probability algorithm model is a deep generative model, which is composed of multiple restricted Boltzmann machines (RBMs) stacked together, that is, the basis of DBN is multiple RBMs, each RBM has a visible layer and a hidden layer, and each RBM is trained in an unsupervised manner so that it can learn the potential feature distribution of the input data, and the trained RBMs are stacked layer by layer to form a deep DBN structure. The hidden layer output of each RBM will be used as the visible layer input of the next RBM. The DBN probability algorithm model learns the high-level abstract features of the input data through unsupervised pre-training and supervised fine-tuning of multiple RBMs, thereby achieving good results in generation and recognition tasks.

[0111] Through the DBN probability algorithm model, the probability distribution information of the disease category of the current user's colorectal disease can be obtained using family disease information. For example, it can be obtained that the probability of the current user suffering from intestinal polyps is 20%, the probability of suffering from proctitis is 16%, and so on.

[0112] In step S107, the first prediction feature and the second prediction feature are input into the trained neural network model to obtain the risk prediction result. It should be noted that the neural network model shown in this step can be obtained based on the training of the convolutional neural network model. The first prediction feature covers the current user's living, eating, sleeping and behavioral characteristics in the second preset time period, and the second prediction feature covers the probability of the current user suffering from various colorectal diseases at the current stage. After deep learning of the neural network model, the probability evolution of various colorectal diseases can be obtained when the current user maintains the behavior and eating habits of the first prediction feature. This evolution information is also the risk prediction result, that is, the risk probability change information of the current user who may suffer from colorectal diseases in the second preset time period, and the category of colorectal diseases that may be suffered. It should be noted that the output result of this step changes dynamically. For example, the actual characteristic information recorded by the user in the actual life information is that the user likes spicy food, has irregular work and rest, stays up late, and has repeated enteritis. The predicted characteristic information corresponding to the first predicted feature is that the current user still likes spicy food, has irregular work and rest, and stays up late during the second preset time period. The risk prediction result shows a gradually increasing trend in the risk probability change of developing enteritis. The colorectal disease that may be suffered may include cell lesions caused by enteritis in a certain part of the intestine on the basis of enteritis, and then there is a high probability of intestinal cancer in that part of the intestine.

[0113] This embodiment is based on a multi-dimensional colorectal cancer risk prediction method. By combining the user's actual life information and family disease information, it realizes the dynamic prediction of the user's future colorectal cancer risk. This dynamic, multi-dimensional risk prediction result is more accurate and comprehensive than the existing methods, and realizes the dynamic and accurate prediction of colorectal cancer risk, providing users with more targeted health management suggestions, and has strong practical value.

[0114] In some embodiments, the real life information is represented as a first multi-dimensional vector, and the first multi-dimensional vector includes a plurality of first information vectors of different dimensions including:

[0115] The actual life information is grouped in sequence according to the actual diet information, the actual work and rest information, the actual behavior information and the actual gastrointestinal information to obtain the first group, the second group, the third group and the fourth group. The first group is the actual diet information, the second group is the actual work and rest information, the third group is the actual behavior information, and the fourth group is the actual gastrointestinal information.

[0116] Extracting non-numerical information in the first group to obtain first category information, performing one-hot encoding on the first category information to obtain a first encoding vector, and extracting numerical information in the first group to obtain first indicator information, performing a logical operation on the first indicator information to obtain a first feature scalar corresponding to the first encoding vector, and combining the first encoding vector with the first feature scalar to form a first information vector of a first dimension;

[0117] Extracting the non-numerical information in the second group to obtain second category information, performing one-hot encoding on the second category information to obtain a second encoding vector, and extracting the numerical information in the second group to obtain second indicator information, performing a logical operation on the second indicator information to obtain a second feature scalar corresponding to the second encoding vector, and combining the second encoding vector with the second feature scalar to form a first information vector of a second dimension;

[0118] Extracting the non-numerical information in the third group to obtain third category information, performing one-hot encoding on the third category information to obtain a third encoding vector, and extracting the numerical information in the third group to obtain third indicator information, performing a logical operation on the third indicator information to obtain a third feature scalar corresponding to the third encoding vector, and combining the third encoding vector with the third feature scalar to form a first information vector of a third dimension;

[0119] Extracting the non-numerical information in the fourth group to obtain fourth category information, performing one-hot encoding on the fourth category information to obtain a fourth encoding vector, and extracting the numerical information in the fourth group to obtain fourth indicator information, performing a logical operation on the fourth indicator information to obtain a fourth characteristic scalar corresponding to the fourth encoding vector, and combining the fourth encoding vector with the fourth characteristic scalar to form a first information vector of a fourth dimension;

[0120] The first information vector of the first dimension, the first information vector of the second dimension, the first information vector of the third dimension, and the first information vector of the fourth dimension are concatenated to form a first multi-dimensional vector.

[0121] In this embodiment, the actual life information is first grouped in sequence according to the actual dietary information, the actual work and rest information, the actual behavioral information, and the actual gastrointestinal information, corresponding to the first group, the second group, the third group, and the fourth group. Further, the data in each group has a minimum recording unit. For example, when the recording unit is a day, the first group contains the actual dietary information of each day within the first preset time period. Each first information vector in the first group represents the actual dietary information of a day. Similarly, the first information vector in the second group, the first information vector in the third group, and the first information vector in the fourth group can be understood by referring to the first information vector in the first group.

[0122] Furthermore, a specific description is given by taking the generation process of the first information vector in the first group as an example.

[0123] In this embodiment, the actual dietary information includes two types: non-numerical and numerical. The non-numerical category includes dietary content expressed in words, for example, specific dishes or categories to which dishes belong, for example, meat dishes, fish dishes, shellfish dishes, leafy vegetables, root vegetables, etc. The dietary content also includes dietary flavors, such as light breakfast, slightly spicy lunch, and medium spicy dinner. The flavor can be specifically combined with each dish, and the current user's daily dietary content includes: breakfast + light + eggs + soy milk, lunch + boiled beef (beef dish) + rice + slightly spicy, dinner + Shanghai green (leafy green) + tofu fish soup (fish dish) + porridge + light; this content can be combined to form a first category information, and the first category information contains one day's breakfast, lunch and dinner content.

[0124] Further, the first category information is one-hot encoded to obtain a first encoding vector. In this embodiment, one-hot encoding is a method of converting category features into numerical features. Specifically, binary values ​​can be used as the encoding form of one-hot encoding. That is, after the first category information is one-hot encoded, a binary numerical sequence can be formed, and the length of the binary numerical sequence can be appropriately adjusted according to the complexity of the first category information. Optionally, the first category information can be embodied in a matrix form. For example, the first row of the matrix is ​​breakfast, the second row is lunch, and the third row is dinner. The first column is the taste feature, the second column is the meat feature, the third column is the vegetarian feature, and the fourth column is the staple food feature. After one-hot encoding, a first encoding matrix will be formed.

[0125] Numerical information includes information whose attributes are numerical values, such as the actual number of meals, the actual meal duration, the actual meal quality, etc. The actual number of meals includes the number of meals in a single day, the actual meal duration is the duration information of each meal in a single day, and the actual meal quality is the quality of food ingested during a single meal. This refined data can be selected according to actual needs to facilitate improving the subsequent prediction accuracy.

[0126] The numerical information in the first group is extracted to obtain the first indicator information, and the first indicator information is logically operated to obtain the first characteristic scalar corresponding to the first coding vector. It should be noted that the logical operation process of the first indicator information can be converted in the form of a preset conversion rule. For example, a ten-point scoring system is set. If the actual eating time is the most reasonable in a certain section, it can correspond to the qualified score section in the scoring system, which can be 7-10 points. If the actual eating time is too short, it can correspond to the low score section in the scoring system, which can be 1-2 points, and so on.

[0127] The numerical value in the first coding vector is used as the category feature vector, and each category feature vector corresponds to a first feature scalar, thereby forming a complete first information vector. Repeat the above steps for multiple actual dietary information in the first group to obtain multiple first information vectors. In this embodiment, the first information vector of the first group is corresponded to the first information vector of the first dimension.

[0128] Similarly, the specific generation steps of the first information vectors of the second group, the third group, and the fourth group can be understood by referring to the first information vector of the first group. The difference is that the category characteristics in the first information vectors under different categories have different meanings. In the actual work and rest information, the category characteristics of the first information vector refer to the specific characteristics in the work and rest information.

[0129] By concatenating the first information vectors in the first group, the second group, the third group and the fourth group, a first multi-dimensional vector can be obtained.

[0130] This embodiment achieves a comprehensive and detailed analysis of the user's living status by systematically representing the actual life information as a first multidimensional vector. The actual life information is divided into multiple dimensions such as diet, work and rest, behavior, and gastrointestinal, making the organization and processing of data clearer. By extracting and encoding this information, especially performing unique hot encoding on non-numerical information and logical operations on numerical information, the formed feature scalar can accurately reflect the user's living habits and health status. Such a structured representation not only improves the availability and accuracy of the data, but also provides high-quality input features for the subsequent LSTM network model, thereby enhancing the predictive ability and practicality of the LSTM network model.

[0131] See also Figure 2 In some embodiments, the first information vector of multiple dimensions includes multiple first feature points, the LSTM network model includes an encoder module and a decoder module, and the first prediction feature includes multiple second feature points;

[0132] The first multidimensional vector is input into the LSTM network model, and the first prediction feature is obtained including:

[0133] S201, sorting first information vectors of the same dimension according to a time series, and dividing a plurality of first information vectors according to a preset time step to obtain a first data set, wherein the first data set includes a plurality of first feature points;

[0134] S202, inputting the first feature points in the first data set into the encoder module one by one according to a preset time step;

[0135] S203, the encoder module performs encoding operation according to the plurality of first feature points to obtain a first hidden state corresponding to each first feature point, and inputs each first hidden state into the attention weight algorithm model to obtain a first attention weight corresponding to each first hidden state;

[0136] S204, performing weighted summation on the multiple first hidden states according to the first attention weight to obtain a first context vector;

[0137] S205, inputting the first context vector into a decoder module;

[0138] S206, the decoder module generates a second hidden state of a plurality of second feature points divided according to a preset time step according to the first context vector;

[0139] S207, inputting each second hidden state into the attention weight algorithm model to obtain a second attention weight corresponding to each second hidden state;

[0140] S208, performing weighted summation on the plurality of second hidden states according to the second attention weight to obtain a first information feature;

[0141] S209, inputting the first information feature and the plurality of second hidden states into a fully connected layer to obtain a second data set, wherein the second data set includes a plurality of second feature points;

[0142] S210, repeat the above steps until each dimension in the first multidimensional vector has a second data set;

[0143] S211. Perform feature concatenation on multiple second data sets to form a first prediction feature.

[0144] In this embodiment, the first multidimensional vector has four groups, and the characteristic information attributes of each group are completely different. In this embodiment, the time series features of the first information vector in each dimension are extracted and predicted one by one. That is, the prediction results of multiple dimensions are output, and finally the prediction results are spliced ​​to form the first and predicted features.

[0145] In step S201, the first information vectors of the same dimension are sorted according to the time series. The preset time step can be set according to actual needs. In this embodiment, it is preferred to set the preset time step to 1 day, which corresponds to the cycle of dietary data collection for a single day in the first information vector. And multiple first information vectors in the same dimension are segmented in units of days. It can be understood that if the first information vector is a meal as the smallest counting unit, the preset time step can be refined to one hour to ensure that the features of the first information vector are completely segmented. After sorting according to the time series, a first data set is formed. The first data set contains sequentially distributed first information vectors. For ease of expression, the first information vector in the first data set is recorded as the first feature point.

[0146] In step S202, the first feature points in the first data set are input into the encoder module one by one according to the preset time step for layer-by-layer encoding. During this process, the encoder module will learn the regular characteristics of the first feature points in the first data set changing over time to facilitate subsequent decoding prediction.

[0147] In step S203, the encoder module performs encoding operations on the first feature points in sequence according to the input order to obtain the first hidden state corresponding to each first feature point. It can be understood that after obtaining the first hidden state of each first feature point, multiple first hidden states are input into the attention weight algorithm model to distribute the attention weight. The attention weight algorithm can reasonably distribute the weight value of the first hidden state in each preset time step, which can enhance the LSTM's understanding of the input sequence and improve the encoding effect. That is, since the LSTM network model relies on the data of the previous first hidden state as decoding, as the number of iterations of the first hidden state increases, there is a problem of encoding error. By using the attention weight algorithm, multiple first hidden states can be reasonably weighted in the global feature field, thereby reducing the error generated by the LSTM network model in the encoding process. For ease of expression, the attention weight corresponding to the first hidden state is recorded as the first attention weight. It can be understood that the attention weight algorithm model used in step S207 is the same model as the attention weight algorithm model in this step, and the difference between the two steps lies in the difference in input data.

[0148] In step S204, multiple first hidden states are weighted and summed according to the first attention weight to obtain a first context vector. The first context vector will be used as input data of the decoder module in step S205, that is, after being encoded by the encoder module, the first context vector covers the time-varying characteristic information of the first information vector in the current dimension.

[0149] In step S206, the decoder module generates a second hidden state of multiple second feature points in a second preset time period in the future according to the first context vector, and divides the second preset time period in the future according to the preset time step to obtain the second hidden state of the second feature point corresponding to the preset time step. It should be noted that the second feature point can be understood as the prediction information of the same dimension in the first prediction feature. For example, in predicting dietary information, each second feature point can be understood as the dietary information of the current user on a certain day in the future.

[0150] In step S207, similar to the first hidden state, in this step the predicted second hidden state is input into the attention weight algorithm model to obtain the second attention weight corresponding to each second hidden state, which can also reduce the error generated by the LSTM network model during the decoding process.

[0151] In step S208, the multiple second hidden states are weighted and summed according to the second attention weight to obtain the first information feature, which aggregates the feature information of multiple second feature points. The first information feature can be used as an input parameter of the fully connected layer to make a more accurate prediction of the second hidden state, thereby improving the accuracy of subsequent prediction results.

[0152] In step S209, the first information feature and the second hidden state are input into the fully connected layer, which is a structure in the LSTM network model. The LSTM network model can reasonably estimate the specific information of the second feature point based on the first information feature and the second hidden feature, thereby obtaining the prediction information of the second feature point. For ease of expression, the output result of the fully connected layer is collected in the second data set, which corresponds to the first data set. The second data set contains the prediction information of the second feature points arranged in the order of the second preset time period, that is, the prediction information of the current dimension in the second preset time period.

[0153] Steps S201 to S209 show the process of inputting a first data set in one dimension into the LSTM network model and finally outputting a second data set. On this basis, when the first multidimensional vector has multiple dimensions, the aforementioned steps are repeated to generate second data sets in multiple dimensions, as shown in step S210, until each dimension of the data in the first multidimensional vector corresponds to a second data set.

[0154] In step S211, features of multiple second data sets are concatenated to form first prediction features.

[0155] Further, in some embodiments, the encoder module includes a plurality of first LSTM units, the first LSTM unit includes a first input gate, a first forget gate, and a first output gate, the decoder module includes a plurality of second LSTM units, the second LSTM unit includes a second input gate, a second forget gate, and a second output gate;

[0156] The encoder module performs encoding operation according to the first feature point currently input, and obtains the first hidden state corresponding to each first feature point including:

[0157] The following steps are performed for each first feature point input to the encoder module one by one:

[0158] Reading a first feature point corresponding to a current preset time step, concatenating a first initial input value of the first feature point with a first hidden state of the first feature point corresponding to a previous preset time step, to obtain a first final input value of the first feature point corresponding to the current preset time step;

[0159] Calculating an activation value of a first input gate, an activation value of a first forget gate, and an activation value of a first output gate according to the first final input value;

[0160] Update the first cell state of the first feature point corresponding to the current preset time step;

[0161] Calculate a first hidden state of a first feature point corresponding to a current preset time step according to an activation value of a first input gate, an activation value of a first forget gate, an activation value of a first output gate, and a first cell state;

[0162] The decoder module generates a second hidden state of a plurality of second feature points divided according to a preset time step according to the first context vector, including:

[0163] The first context vector is used as the second hidden state of the first second feature point, and the second cell state of the first second feature point is initialized to zero;

[0164] The following steps are performed for each second feature point input to the decoder module one by one:

[0165] Reading an output value of the second feature information corresponding to the previous preset time step as a second initial input value of the second feature point corresponding to the current preset time step, and concatenating the second initial input value with the first context vector to obtain a second final input value of the second feature point corresponding to the current preset time step;

[0166] Calculating an activation value of a second input gate, an activation value of a second forget gate, and an activation value of a second output gate according to the second final input value;

[0167] Updating the second cell state of the second feature point corresponding to the current preset time step;

[0168] The second hidden state of the second feature point corresponding to the current preset time step is calculated according to the activation value of the second input gate, the activation value of the second forget gate, the activation value of the second output gate, and the second cell state.

[0169] In this embodiment, the specific operation steps of the encoder module and the decoder module are shown. Specifically, the encoder module contains multiple first LSTM units, each of which includes a first input gate, a first forget gate and a first output gate. Multiple first LSTM units are in a stacked form, and the first LSTM unit operation is performed on each first feature point input one by one, and the first hidden state of the previous first feature point is spliced ​​with the first initial input value of the current first feature point as the first final input value, and then the first final input value is used as the calculation of the activation value of the first input gate, the first output gate and the first forget gate, and the first cell state of the current first feature point is updated, and the first hidden state of the current first feature point is further calculated based on the above data.

[0170] That is, the encoder module relies on the first hidden state of the previous first feature point to calculate the first hidden state of the current first feature point during the encoding process. This method can well retain the characteristics of each first feature point and extract features in repeated iterative calculations to obtain the final first context vector.

[0171] Similarly, the decoder module contains multiple second LSTM units, each of which includes a second input gate, a second forget gate, and a second output gate. The multiple second LSTM units are in a stacked form, and correspondingly, the encoder module and the decoder module are two opposite modules. The encoder module uses the first feature point to obtain the first hidden feature, and the decoder module uses multiple second hidden features to calculate the second feature point; in this process, the input data of the decoder module is the second hidden feature, and the output data of the decoder module is the second feature point, that is, the prediction information. In this embodiment, the first context vector is used as the second hidden state of the first second feature point, and the first second feature point is the second feature point at the first position when the second preset time period is divided according to the preset time step; the second cell state of the first second feature point is initialized to 0.

[0172] After concatenating the first context vector with the second initial input value of the current second feature point, the second final input value is used as the second input value, and then the second final input value is used to calculate the activation values ​​of the second input gate, the second output gate and the second forget gate, and the second cell state of the current second feature point is updated, and the second hidden state of the current second feature point is further calculated based on the above data.

[0173] That is, the decoder module relies on the first context vector to calculate the second hidden state of multiple subsequent second feature points during the decoding process. This method can well retain the characteristics of each second feature point, and at the same time apply it to the feature information covered in the first context, and extract features in a repeated iterative calculation process to obtain the final first information feature, which aggregates the feature information of multiple second hidden states.

[0174] This embodiment effectively realizes the feature extraction and prediction of multi-dimensional time series data by constructing an LSTM network model and combining the attention mechanism. By gradually encoding and decoding the first information vector of each dimension, the corresponding prediction information of the second feature point is generated, which can accurately capture the changing pattern of the user's life information over time. The introduction of the attention mechanism not only improves the LSTM network model's understanding of the input sequence, but also reduces the errors in the encoding and decoding process, thereby improving the accuracy and reliability of the prediction information. Finally, by splicing the prediction information under multiple dimensions, a comprehensive first prediction feature is formed, which is convenient for the analysis and calculation of the subsequent neural network model.

[0175] See also Figure 3 In some embodiments, the family disease information is represented as a second multidimensional vector, and the second multidimensional vector includes a plurality of second information vectors of different dimensions including:

[0176] Dividing the plurality of colorectal disease information in the family disease information by family members to obtain disease information of each family member;

[0177] Perform the following steps for each family member's medical condition information:

[0178] S301, extracting non-numerical information from the family member's symptom information to obtain first symptom information;

[0179] S302, performing one-hot encoding on the first symptom information to obtain a first symptom vector; and extracting numerical information from the family member's symptom information to obtain second symptom information;

[0180] S303, performing a logical operation on the second symptom information to obtain a first symptom scalar corresponding to the first symptom vector;

[0181] S304, combining the first symptom vector and the first symptom scalar to form a second information vector;

[0182] S305: Concatenate multiple second information vectors to form a second multi-dimensional vector.

[0183] In this embodiment, the family disease information is divided according to the family members, and the family disease information of each family member is obtained. In step S301, the non-numerical information in the family disease information can be understood by referring to the non-numerical information in the actual diet information, and can be specifically the symptom description content, for example, abdominal pain + accompanied by blood in the stool + enteritis, to obtain the first disease information. It is understandable that when a family member has multiple historical diseases, the first disease information can be embodied in the form of a group.

[0184] In step S302, the first symptom information is one-hot encoded to digitize the first symptom information, and the numerical information in the family member's symptom information is extracted to obtain the second symptom information.

[0185] In step S303, the logical operation of the second symptom information may be normalized, specifically, Min-Max normalization or Z-score normalization, to convert it into a numerical value of a unified evaluation standard.

[0186] In step S304, the first symptom vector and the first symptom scalar may be multiplied in the same dimension to obtain a second information vector.

[0187] In step S305, the above steps may be repeated until the second information vectors of multiple family members are converted, and finally a second multi-dimensional vector is formed.

[0188] This embodiment makes full use of the non-numerical and numerical features in the symptom information by structuring the family symptom information into a second multidimensional vector, converting the symptom information of each family member into a second information vector containing symptom categories and numerical features, which can more comprehensively describe multiple aspects of the symptom, effectively digitize the original unstructured symptom information, and greatly improve the usability and operability of the data. It makes full use of the complementarity of different types of information and enhances the comprehensive expression ability of the symptom status. The constructed second multidimensional vector has good structural characteristics, which is convenient for the accuracy of subsequent operations.

[0189] In some embodiments, the DBN probability algorithm model includes multiple RBM layers stacked from top to bottom, each RBM layer includes visible units and hidden units, and the multiple RBM layers are numbered in sequence from top to bottom in the stacking order, the topmost RBM layer is recorded as the first RBM layer, and the bottommost RBM layer is recorded as the Nth RBM layer. The second multidimensional vector is input into the DBN probability algorithm model, and the second prediction feature is obtained, including:

[0190] The second multidimensional vector is used as the input feature of the visible unit of the first RBM layer;

[0191] The first RBM layer generates the feature code of the hidden unit of the first RBM layer according to the second multidimensional vector, which is recorded as the first feature code and expressed by formula (1). Formula (1) is as follows:

[0192] ;

[0193] In formula (1), is the first feature code, is the activation function of the first RBM layer, is the weight matrix of the first RBM layer, is the bias vector of the first RBM layer;

[0194] The first feature code is used as the input feature of the visible unit of the second RBM layer to generate the second feature code of the second RBM layer, which is expressed by formula (2). Formula (2) is as follows:

[0195] ;

[0196] In formula (2), For the The input features of an RBM are Encode the features of the previous RBM layer, , For the The activation function of the RBM layer is For the The weight matrix of each RBM layer is For the The bias vector of each RBM layer;

[0197] Repeat the above steps until the Nth RBM layer is reached, and the Nth feature code of the Nth RBM layer is obtained. The Nth feature code is used as the second prediction feature and substituted into formula (2) The second prediction feature is obtained, which is expressed as formula (3). Formula (3) is as follows:

[0198] ;

[0199] In formula (3), is the second prediction feature, For the The input features of an RBM are: For the The activation function of the RBM layer is For the The weight matrix of each RBM layer is For the The bias vector of the RBM layer.

[0200] In this embodiment, the RBM layers are stacked in order, and the data output of the previous RBM layer will be used as the data input of the next RBM layer. Finally, the features extracted by each RBM layer can be aggregated to form the second prediction feature, that is, the high-level feature representation.

[0201] This embodiment effectively realizes the process of extracting high-level features from the second multi-dimensional vector layer by layer by constructing a DBN probabilistic algorithm model, significantly improving the richness and complexity of the feature representation of the second multi-dimensional vector. Through layer-by-layer learning of multi-layer stacked restricted Boltzmann machines (RBMs), the DBN probabilistic algorithm model can capture the deep-level patterns and potential structures in the second multi-dimensional vector, thereby generating a second prediction feature with higher discriminative ability. This feature extraction method not only enhances the DBN probabilistic algorithm model's ability to understand complex data, but also improves the accuracy and reliability of subsequent prediction tasks.

[0202] See also Figure 4 and Figure 5 In some embodiments, the method further comprises:

[0203] S401, generating a first disease curve of a plurality of colorectal diseases within a second preset time period according to the risk prediction result and the first prediction feature, wherein the first disease curve is a disease probability change curve predicted by the current user; and screening a plurality of first individual vectors corresponding to a first group whose vector similarity with the first multidimensional vector is placed in the same vector similarity range in the database;

[0204] S402, obtaining a second disease curve corresponding to each first body vector;

[0205] S403, matching the slope similarities of the first disease curve and the second disease curve one by one, dividing the individuals of the second disease curve in the same slope similarity range to obtain a second group, and recording the first body vector in the second group as a second body vector;

[0206] S404, obtaining a second intervention curve corresponding to each second individual vector, where the second intervention curve is an actual disease probability change curve of the second group after receiving the corresponding second intervention scheme;

[0207] S405, generating a primary intervention plan corresponding to the current user according to the multiple second intervention curves;

[0208] S406: Obtain fine-tuning information, where the fine-tuning information is generated according to manual input by the user, and the primary intervention plan is adjusted according to the fine-tuning information to obtain a final intervention plan for the current user.

[0209] In step S401, a first disease curve of colorectal diseases with the first prediction feature as an independent variable can be generated according to the risk prediction result and the first prediction feature. It should be noted that there can be multiple first disease curves, each corresponding to a disease category. Further, multiple first body vectors corresponding to the first group whose similarity with the first multidimensional vector is placed in the same vector similarity range are screened out from the database, that is, the first body vectors placed in the same vector similarity range are integrated to form a first group, and the first body vector is the first multidimensional vector corresponding to other patients in the database. Furthermore, the cosine similarity algorithm can be used to calculate the similarity between the first multidimensional vector and the first body vector.

[0210] In step S402, the second disease curve corresponding to each first body vector is obtained. That is, in the database, the first body vector also corresponds to a risk prediction result, and the second disease curve of the patient can also be generated according to the risk prediction result.

[0211] In step S403, the slope similarities of the first disease curve and the second disease curve are matched one by one. It should be noted that this slope similarity can be a slope similarity comparison of each preset time step, that is, the first disease curve is divided into a plurality of first disease segments according to the preset time step, and further, the second disease curve is divided into a plurality of second disease segments according to the same preset time step, and the similarity of the slope of the first disease segment and the slope of the second disease segment is judged one by one according to the arrangement order of the preset time step, and finally the slope similarity judgment results of the plurality of segments are integrated into a similarity probability.

[0212] In step S403, individuals of the second disease curve whose slope similarity is in the same slope similarity range as the first disease curve are separated from the first group and recorded as the second group, and the first body vector in the second group is recorded as the second body vector for easy distinction.

[0213] In step S404, the second intervention curve corresponding to each second body vector is obtained. It should be noted that the second intervention curve is the actual disease probability change curve of the second group after receiving the corresponding second intervention plan; that is, the patients in the database are all patients who have received lifestyle intervention after consultation, and each patient has an intervention plan. For easy distinction, the intervention plan for patients in the first group is recorded as the first intervention plan, and the intervention plan for patients in the second group is recorded as the second intervention plan. After executing the second intervention plan, the actual disease probability change curve of patients in the second group is different from the second disease curve, which shows that the second intervention plan has a better disease prevention effect. Figure 5 As shown, Figure 5 The solid line in the middle is the second disease curve, and the dotted line is the second intervention curve.

[0214] In step S405, a primary intervention plan corresponding to the current user is generated according to the plurality of second intervention curves.

[0215] For further information, see Figure 5 and Figure 6 In some embodiments, generating a primary intervention plan corresponding to the current user according to the plurality of second intervention curves includes:

[0216] Calculating the slope difference between each second intervention curve and its corresponding second disease curve;

[0217] Arrange the multiple slope differences from large to small, and select the second intervention plan corresponding to the second intervention curve with the largest slope difference as the primary intervention plan.

[0218] The calculation of the slope difference may specifically be: dividing the second intervention curve into a plurality of second intervention segments according to a preset time step, and calculating the slope difference between the second intervention segment and the second diseased segment one by one according to the arrangement order of the preset time steps, such as Figure 5 and Figure 6 As shown, we can see that Figure 5 In the changing trend, the slopes of the second intervention curve and the second disease curve gradually become the same, but in the early stage, the slope of the second intervention curve is smaller than the slope of the second disease curve. Figure 6 A change state of the slope difference is shown in FIG. By calculating multiple second intervention curves and multiple second disease curves, a change trend graph of the slope difference of each patient can be obtained. The change in the slope difference represents the effective prevention trend of the second intervention plan after implementation. In this embodiment, by calculating the slope difference between the second intervention curve and the second disease curve, the second intervention plan can be screened, and the second intervention plan with the most significant effect can be selected as the primary intervention plan.

[0219] In step S406, fine-tuning information is obtained. It should be noted that the fine-tuning information is manually input by the user. For example, some users have dietary taboos due to religious reasons, so they need to fine-tune some individual ingredients in the primary intervention plan. Some users have special work schedules, such as two-shift work schedules, so they need to fine-tune the work and rest schedule in the primary intervention plan. And so on, finally get the final intervention plan that matches the current user. It can be understood that the intervention plan covers many aspects of intervention indicators in the predicted life information, such as diet planning, work and rest planning, exercise planning, etc., so as to achieve early prevention and avoidance of colorectal cancer.

[0220] This embodiment uses the risk prediction result and the first prediction feature to generate the colorectal disease prediction curve of the current user, and compares and analyzes it with the first body vector of other patients in the database to find a first group similar to it. For the similar population screened out, the similarity between the first disease curve and the second disease curve is further analyzed to obtain the second group, and then the second intervention curve after the actual intervention in the second group is obtained, from which the most effective intervention plan is mined and used as the primary intervention plan for the current user. This can not only more accurately predict the disease risk of the current user, but also draw on the intervention experience of similar patients. This embodiment allows the user to fine-tune the primary intervention plan according to the fine-tuning information generated by the user's own characteristics, further optimize the intervention plan, ensure that it is more in line with the current user's living conditions and needs, and can greatly improve the pertinence and feasibility of the intervention plan, and improve the user's active participation and the executability of the plan.

[0221] In a second aspect, this embodiment further provides a computer-readable storage medium on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described in the first aspect is implemented.

[0222] The computer program involved in this embodiment can be stored in a computer device readable storage medium, which includes but is not limited to a disk, a tape, a magnetic card, a floppy disk, a flash memory, an optical disk, an optical card, a read-only memory (ROM), a random access memory (RAM), an erasable programmable ROM (EPROM) and an electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical or chemical structures that can achieve the same or equivalent functions as the storage media listed above, such as DNA, RNA, protein and other units with information storage capabilities. In a specific embodiment, the storage medium involved can be one of the above-mentioned media types, or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiment can be stored in a single medium in a centralized manner, or it can be stored in multiple media in a distributed manner. The memory containing the computer device readable storage medium can be a non-volatile memory or a random access memory. These computer device readable storage media can be built into the device, or they can be connected to the device involved in the embodiment as an external device or a part of an external device. In some embodiments, a memory with a computer device readable storage medium is deployed locally; in other embodiments, a solution of deploying the memory away from the processor may also be adopted, such as a network attached memory accessed via an RF circuit or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), etc., or a suitable combination thereof, as long as the computer device can access the memory. In addition, the computer program involved in the embodiment may be stored in plaintext / ciphertext form, or may be designed as training data, which may be integrated and reorganized through model training and implicitly stored in the parameter state of a deep neural network or other machine learning model.

[0223] In a third aspect, this embodiment further provides an electronic device, comprising a memory and a processor, wherein the memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method described in the first aspect.

[0224] The processor described in this embodiment can be implemented by hardware, firmware, software or a combination thereof, and can use a circuit, a single or multiple application-specific integrated circuits (Application Specific Integrated Circuit, ASIC), a digital signal processor (Digital Signal Processor, DSP), a digital signal processing device (Digital Signal Processing Device, DSPD), a programmable logic device (Programmable Logic Device, PLD), a field programmable gate array (Field Programmable Gate Array, FPGA), a central processing unit (Central Processing Unit, CPU), a controller, a microcontroller, at least one of a microprocessor, and also includes other physical, biological or chemical structures that can achieve similar or equivalent functions to the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of the present application, or any combination of the steps mentioned therein.

[0225] The risk prediction method shown in the above embodiment uses the LSTM network model to predict the user's life status changes in the future period of time according to the user's current actual life information (diet, work and rest, behavior, gastrointestinal, etc.), which provides important basic data for subsequent colorectal cancer risk prediction. In addition, the risk prediction method uses the DBN probability algorithm model to integrate the colorectal disease information of the user's family, and generates the prediction characteristics of the probability distribution of the user's own colorectal disease, which can more comprehensively evaluate the user's genetic risk factors. The above two prediction features are input into the trained neural network model, which can not only obtain the risk probability changes of the user's future colorectal cancer, but also predict the specific colorectal disease types that may be suffered. This technical solution is based on a multi-dimensional colorectal cancer risk prediction method. By combining the user's actual life information and family disease information, it realizes the dynamic prediction of the user's future colorectal cancer risk. This dynamic, multi-dimensional risk prediction result is more accurate and comprehensive than the existing method, and realizes the dynamic and accurate prediction of colorectal cancer risk, providing users with more targeted health management suggestions, and has a strong practical value.

[0226] Finally, it should be noted that although the above embodiments have been described in the specification and drawings of this application, this does not limit the scope of patent protection of this application. All technical solutions generated by replacing or modifying equivalent structures or equivalent processes based on the essential concept of this application using the contents recorded in the specification and drawings of this application, as well as directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, are included in the scope of patent protection of this application.

Claims

1. A multi-dimensional colorectal cancer risk prediction method, characterized in that: include: Acquire actual life information of the current user within a first preset time period, wherein the actual life information includes actual diet information, actual work and rest information, actual behavior information, and actual gastrointestinal information; Representing the real life information as a first multi-dimensional vector, wherein the first multi-dimensional vector includes a plurality of first information vectors of different dimensions; Inputting the first multidimensional vector into an LSTM network model to obtain a first prediction feature, the LSTM network model is obtained by fusion through an attention mechanism, and the first prediction feature is the predicted life information of the current user within a second preset time period; Acquire family disease information of the current user, wherein the family disease information includes colorectal disease information of multiple family members; Representing the family disease information as a second multi-dimensional vector, wherein the second multi-dimensional vector includes a plurality of second information vectors of different dimensions; Inputting the second multidimensional vector into the DBN probability algorithm model to obtain a second prediction feature, where the second prediction feature is probability distribution information of the disease category of the colorectal disease of the current user; Inputting the first prediction feature and the second prediction feature into a trained neural network model to obtain a risk prediction result, wherein the risk prediction result is risk probability change information of the current user who may suffer from a colorectal disease within a second preset time period and the disease category of the colorectal disease that the user may suffer from; generating a first disease curve of a plurality of colorectal diseases within a second preset time period according to the risk prediction result and the first prediction feature, Selecting a plurality of first body vectors corresponding to a first group whose vector similarity is in the same vector similarity range as the first multidimensional vector; Obtaining a second disease curve corresponding to each first body vector; Matching the slope similarity of the first disease curve and the second disease curve one by one to obtain a second group, and recording the first body vector in the second group as a second body vector; Obtaining a second intervention curve corresponding to the second body vector to generate a primary intervention plan corresponding to the current user; The fine-tuning information is obtained, and the primary intervention plan is adjusted according to the fine-tuning information to obtain the final intervention plan for the current user.

2. The multi-dimensional colorectal cancer risk prediction method according to claim 1, characterized in that: The real life information is represented as a first multi-dimensional vector, wherein the first multi-dimensional vector includes a plurality of first information vectors of different dimensions including: The actual life information is grouped in sequence according to actual dietary information, actual work and rest information, actual behavior information, and actual gastrointestinal information to obtain a first group, a second group, a third group, and a fourth group, wherein the first group is actual dietary information, the second group is actual work and rest information, the third group is actual behavior information, and the fourth group is actual gastrointestinal information; Extracting non-numerical information in the first group to obtain first category information, performing one-hot encoding on the first category information to obtain a first encoding vector, and extracting numerical information in the first group to obtain first indicator information, performing a logical operation on the first indicator information to obtain a first feature scalar corresponding to the first encoding vector, and combining the first encoding vector with the first feature scalar to form a first information vector of a first dimension; Extracting the non-numerical information in the second group to obtain second category information, performing one-hot encoding on the second category information to obtain a second encoding vector, and extracting the numerical information in the second group to obtain second indicator information, performing a logical operation on the second indicator information to obtain a second feature scalar corresponding to the second encoding vector, and combining the second encoding vector with the second feature scalar to form a first information vector of a second dimension; Extracting non-numerical information in the third group to obtain third category information, performing one-hot encoding on the third category information to obtain a third encoding vector, and extracting numerical information in the third group to obtain third indicator information, performing a logical operation on the third indicator information to obtain a third feature scalar corresponding to the third encoding vector, and combining the third encoding vector with the third feature scalar to form a first information vector of a third dimension; Extracting non-numerical information in the fourth group to obtain fourth category information, performing one-hot encoding on the fourth category information to obtain a fourth encoding vector, and extracting numerical information in the fourth group to obtain fourth indicator information, performing a logical operation on the fourth indicator information to obtain a fourth characteristic scalar corresponding to the fourth encoding vector, and combining the fourth encoding vector with the fourth characteristic scalar to form a first information vector of a fourth dimension; The first information vector of the first dimension, the first information vector of the second dimension, the first information vector of the third dimension, and the first information vector of the fourth dimension are concatenated to form the first multidimensional vector.

3. The multi-dimensional colorectal cancer risk prediction method according to claim 1, characterized in that: The first information vector of multiple dimensions includes multiple first feature points, the LSTM network model includes an encoder module and a decoder module, and the first prediction feature includes multiple second feature points; The first multidimensional vector is input into the LSTM network model to obtain the first prediction feature including: Arrange multiple first feature points in the same first information vector according to a time series, and segment the multiple first feature points according to a preset time step to obtain a first data set; Inputting the first feature points in the first data set into the encoder module one by one according to a preset time step; The encoder module performs encoding operation according to the plurality of first feature points to obtain a first hidden state corresponding to each first feature point, and inputs each first hidden state into the attention weight algorithm model to obtain a first attention weight corresponding to each first hidden state; Performing weighted summation on the plurality of first hidden states according to the first attention weight to obtain a first context vector; inputting the first context vector into the decoder module; The decoder module generates a second hidden state of a plurality of second feature points divided according to a preset time step according to the first context vector; Input each second hidden state into the attention weight algorithm model to obtain a second attention weight corresponding to each second hidden state; Performing weighted summation on the plurality of second hidden states according to the second attention weight to obtain a first information feature; Inputting the first information feature and the plurality of the second hidden states into a fully connected layer to obtain a second data set, wherein the second data set includes the plurality of the second feature points; Repeat the above steps until each dimension in the first multidimensional vector has a second data set; Perform feature concatenation on multiple second data sets to form the first prediction feature.

4. The multi-dimensional colorectal cancer risk prediction method according to claim 3, characterized in that: The encoder module includes a plurality of first LSTM units, each of which includes a first input gate, a first forget gate, and a first output gate; the decoder module includes a plurality of second LSTM units, each of which includes a second input gate, a second forget gate, and a second output gate; The encoder module performs encoding operation according to the first feature point currently input, and obtains the first hidden state corresponding to each first feature point, including: The following steps are performed for each of the first feature points input to the encoder module one by one: Reading the first feature point corresponding to the current preset time step, concatenating the first initial input value of the first feature point with the first hidden state of the first feature point corresponding to the previous preset time step to obtain a first final input value of the first feature point corresponding to the current preset time step; Calculating an activation value of the first input gate, an activation value of the first forget gate, and an activation value of the first output gate according to the first final input value; Update the first cell state of the first feature point corresponding to the current preset time step; Calculating a first hidden state of the first feature point corresponding to a current preset time step according to an activation value of the first input gate, an activation value of the first forget gate, an activation value of the first output gate, and the first cell state; The decoder module generates a second hidden state of a plurality of second feature points divided according to a preset time step according to the first context vector, comprising: The first context vector is used as a second hidden state of the first second feature point, and the second cell state of the first second feature point is initialized to zero; For each of the second feature points input to the decoder module one by one, the following steps are performed: Reading an output value of the second feature information corresponding to the previous preset time step as a second initial input value of the second feature point corresponding to the current preset time step, and concatenating the second initial input value with the first context vector to obtain a second final input value of the second feature point corresponding to the current preset time step; Calculating an activation value of the second input gate, an activation value of the second forget gate, and an activation value of the second output gate according to the second final input value; Updating the second cell state of the second feature point corresponding to the current preset time step; The second hidden state of the second feature point corresponding to the current preset time step is calculated according to the activation value of the second input gate, the activation value of the second forget gate, the activation value of the second output gate, and the second cell state.

5. The multi-dimensional colorectal cancer risk prediction method according to claim 1, characterized in that: The family disease information is represented as a second multidimensional vector, wherein the second multidimensional vector includes a plurality of second information vectors of different dimensions including: Dividing the plurality of colorectal disease information in the family disease information by family members to obtain disease information of each family member; For each family member's medical condition information, perform the following steps: Extracting non-numerical information from the family member's symptom information to obtain first symptom information; Performing one-hot encoding on the first symptom information to obtain a first symptom vector; and extracting the numerical information from the family member's symptom information to obtain second symptom information; Performing a logical operation on the second symptom information to obtain a first symptom scalar corresponding to the first symptom vector; combining the first symptom vector and the first symptom scalar to form the second information vector; A plurality of second information vectors are concatenated to form the second multi-dimensional vector.

6. The multi-dimensional colorectal cancer risk prediction method according to claim 1, characterized in that: The DBN probability algorithm model includes multiple RBM layers stacked from top to bottom, each of the RBM layers includes visible units and hidden units, and the multiple RBM layers are numbered in sequence from top to bottom in the stacking order, the topmost RBM layer is recorded as the first RBM layer, and the bottommost RBM layer is recorded as the Nth RBM layer. The second multidimensional vector is input into the DBN probability algorithm model, and the second prediction feature is obtained, including: Using the second multidimensional vector as the input feature of the visible unit of the first RBM layer; The first RBM layer generates a feature code of a hidden unit of the first RBM layer according to the second multidimensional vector, which is recorded as the first feature code and is expressed by formula (1). Formula (1) is as follows: ； In formula (1), is the first feature code, is the activation function of the first RBM layer, is the weight matrix of the first RBM layer, is the bias vector of the first RBM layer; The first feature code is used as the input feature of the visible unit of the second RBM layer to generate the second feature code of the second RBM layer, which is expressed by formula (2). Formula (2) is as follows: ； In formula (2), For the The input features of an RBM are Encode the features of the previous RBM layer, , For the The activation function of the RBM layer is For the The weight matrix of each RBM layer is For the The bias vector of each RBM layer; Repeat the above steps until the Nth RBM layer is reached, and the Nth feature code of the Nth RBM layer is obtained. The Nth feature code is used as the second prediction feature, and the formula (2) is substituted into The second prediction feature is obtained, which is expressed as formula (3). Formula (3) is as follows: ； In formula (3), is the second prediction feature, For the The input features of an RBM are: For the The activation function of the RBM layer is For the The weight matrix of each RBM layer is For the The bias vector of the RBM layer.

7. The multi-dimensional colorectal cancer risk prediction method according to claim 1, characterized in that: The first disease curve is a disease probability change curve predicted by the current user; The second intervention curve is a curve of actual disease probability change after the second groups receive their respective corresponding second intervention plans.

8. The multi-dimensional colorectal cancer risk prediction method according to claim 1, characterized in that: Generating a primary intervention plan corresponding to the current user according to the plurality of the second intervention curves includes: Calculating the slope difference between each second intervention curve and its corresponding second disease curve; Arrange the plurality of slope differences from large to small, and select the second intervention plan corresponding to the second intervention curve with the largest slope difference as the primary intervention plan.

9. A computer-readable storage medium storing computer program instructions, characterized in that: The computer program instructions implement the method of any one of claims 1 to 8 when executed by a processor.

10. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method according to any one of claims 1 to 8.

Images