Disease risk prediction method and device, equipment, storage medium and program product

Abstract

The embodiment of the invention provides a disease risk prediction method and device, equipment, a storage medium and a program product. The method comprises the following steps: acquiring historical physical examination information of a target object; inputting the historical examination information of the target object into a disease risk prediction model to obtain a disease risk prediction result corresponding to the target object; wherein the disease risk prediction model is obtained based on training of a time-dependent loss function, and the time-dependent loss function is determined based on a weight adjustment factor related to a time step length and a smoothness regular term. The accuracy of disease risk prediction can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a disease risk prediction method, apparatus, device, storage medium, and program product. Background Art

[0002] The brain is the most important central organ in humans, playing a vital role in maintaining normal physiological functions and mental activity. However, brain diseases have gradually become a serious problem plaguing human health. For example, enlarged peripheral vascular spaces, cerebral microbleeds, and microinfarcts often take months or even years to develop. During this stage, the human body usually has no obvious abnormal symptoms and therefore is often overlooked.

[0003] Currently, a comprehensive assessment of brain disease typically requires an MRI scan, which is expensive and time-consuming, making it unsuitable for routine physical examinations. Therefore, existing technologies typically rely on doctors' medical experience to predict brain disease risk and provide timely warnings. However, due to the significant individual differences in expertise among doctors, the accuracy of these predictions cannot be guaranteed. Summary of the Invention

[0004] Various aspects of the present application provide a disease risk prediction method, apparatus, device, storage medium, and program product to improve the accuracy of disease risk prediction.

[0005] The present invention provides a method for predicting disease risk, including:

[0006] Obtain historical physical examination information of the target subject;

[0007] Inputting the historical physical examination information of the target subject into the disease risk prediction model to obtain the disease risk prediction result corresponding to the target subject;

[0008] Among them, the disease risk prediction model is trained based on a time-related loss function, and the time-related loss function is determined based on a weight adjustment factor related to the time step and a smoothness regularization term.

[0009] In an optional embodiment, the training process of the disease risk prediction model is as follows:

[0010] Acquire training sample data corresponding to the disease risk prediction model and supervision information corresponding to the training sample data;

[0011] Obtaining a plurality of teacher models pre-trained using the training sample data and the supervision information;

[0012] Taking the disease risk prediction model as a student model, inputting the training sample data into the multiple teacher models and the disease risk prediction model respectively, obtaining first probability distribution prediction results output by the multiple teacher models and second probability distribution prediction results output by the disease risk prediction model;

[0013] Determining a target loss function based on the first probability distribution prediction result and the second probability distribution prediction result;

[0014] The parameters of the disease risk prediction model are adjusted according to the loss function.

[0015] In an optional embodiment, the teacher model and the student model are machine learning models that integrate feature engineering, and the feature engineering is used to perform feature selection, feature extraction and feature construction on the input of the machine learning model.

[0016] In an optional embodiment, the features acquired based on the feature engineering include: time series features, statistical features, and change trend features;

[0017] The time series feature is used to reflect the correlation between the various physical examination indicators of the historical physical examination information in the time series;

[0018] The statistical features are used to reflect the overall distribution and dispersion of each physical examination indicator in the historical physical examination information;

[0019] The change trend feature is used to reflect the change trend of each physical examination indicator in the historical physical examination information within the set time period.

[0020] In an optional embodiment, determining a target loss function based on the first probability distribution prediction result and the second probability distribution prediction result includes:

[0021] Determining the cross entropy loss function based on the first probability distribution prediction result and the second probability distribution prediction result;

[0022] Determining the time-dependent loss function based on the first probability distribution prediction result, the second probability distribution prediction result, the weight adjustment factor related to the time step, and the smoothness regularization term;

[0023] A target loss function is determined according to the cross entropy loss function and the time-dependent loss function.

[0024] In an optional embodiment, determining a target loss function according to the cross entropy loss function and the time-dependent loss function includes:

[0025] A weighted sum is performed on the cross entropy loss function and the time-dependent loss function to obtain a target loss function.

[0026] In an optional embodiment, the method further includes:

[0027] performing a performance evaluation on the student model;

[0028] Based on the results of the performance evaluation, the student model is optimized.

[0029] The present invention provides a disease risk prediction device, comprising:

[0030] An acquisition module is used to obtain the historical physical examination information of the target object;

[0031] A prediction module is used to input the historical physical examination information of the target subject into the disease risk prediction model to obtain the disease risk prediction result corresponding to the target subject;

[0032] Among them, the disease risk prediction model is trained based on a time-related loss function, and the time-related loss function is determined based on a weight adjustment factor related to the time step and a smoothness regularization term.

[0033] An embodiment of the present application provides an electronic device, comprising: a memory, a processor, and a communication interface; wherein the memory stores executable code, and when the executable code is executed by the processor, the processor executes the above method.

[0034] An embodiment of the present application provides a non-transitory machine-readable storage medium having executable code stored thereon. When the executable code is executed by a processor of an electronic device, the processor is caused to perform the above method.

[0035] An embodiment of the present application provides a computer program product, which includes: a computer program, which, when executed by a processor of an electronic device, causes the processor to perform the above method.

[0036] In the embodiment of the present application, since the disease risk prediction model is obtained by training based on a time-dependent loss function, and the time-dependent loss function is determined based on a weight adjustment factor related to the time step, and a smoothness regularization term. Therefore, the disease risk prediction model in the embodiment of the present application can make better use of the time series information of the historical physical examination information to improve the accuracy of disease risk prediction. At the same time, it can learn a smoother and more interpretable prediction sequence, which is beneficial to reduce the complexity of the disease risk prediction model while ensuring the prediction performance, making it more suitable for actual clinical scenarios. In actual application, it is only necessary to input the historical physical examination information of the target object into the disease risk prediction model to obtain accurate disease risk prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0038] Figure 1 A flowchart of a disease risk prediction method provided in an exemplary embodiment of the present application;

[0039] Figure 2 A schematic diagram of the training process of a disease risk prediction model provided in an exemplary embodiment of the present application;

[0040] Figure 3 A schematic diagram of the structure of a disease risk prediction device provided by an exemplary embodiment of the present application;

[0041] Figure 4 A schematic structural diagram of an electronic device provided as an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0042] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0043] The brain is the most important central organ of the human body and plays a vital role in maintaining normal physiological functions and mental activities. However, brain diseases have gradually become a serious problem that plagues human health. For example, enlarged peripheral vascular spaces, cerebral microbleeds, microinfarction foci, etc. These brain diseases often take months or even years to form. At this stage, the human body usually has no obvious abnormal symptoms, so they are often ignored. In the related art, it is generally based on the medical experience of doctors to predict the risk of brain diseases for users and to promptly issue brain disease risk warnings to users. However, due to the large individual differences in the knowledge expertise of doctors, the accuracy of the prediction results cannot be guaranteed. In view of this, the embodiment of the present application provides a disease risk prediction method.

[0044] Figure 1 The flowchart of the disease risk prediction method provided in the embodiment of the present application is as follows: Figure 1 As shown, the method includes:

[0045] Step 101: Obtain historical physical examination information of the target object.

[0046] Step 102: Input the historical physical examination information of the target object into the disease risk prediction model to obtain the disease risk prediction result corresponding to the target object, wherein the disease risk prediction model is trained based on the time-related loss function, and the time-related loss function is determined based on the weight adjustment factor related to the time step and the smoothness regularization term.

[0047] In actual applications, it is only necessary to obtain the user's historical physical examination information, and then input the user's historical physical examination information into the trained disease risk prediction model to obtain the disease risk prediction result corresponding to the user. For example, the disease risk prediction result can be the probability of the user suffering from brain disease in the next year, or whether the user will suffer from brain disease in the next year, etc., which is not limited here. It can be determined based on the training samples and supervision information used in the training process of the disease risk prediction model. Among them, the historical physical examination information can be physical examination information within a set time period, for example, the user's blood pressure, cholesterol level, blood sugar, body mass index (BMI), and parameters related to lifestyle habits (such as smoking, drinking, exercise frequency, etc.) within 5 years, which are not listed one by one here.

[0048] It should be noted that since the disease risk prediction model is trained based on a time-dependent loss function, and the time-dependent loss function is determined based on a weight adjustment factor related to the time step and a smoothness regularization term, the disease risk prediction model in the embodiment of the present application can better utilize the time series information of historical physical examination information to improve the accuracy of disease risk prediction. At the same time, it can learn a smoother and more interpretable prediction sequence, which is beneficial to reduce the complexity of the disease risk prediction model while ensuring the prediction performance, making it more suitable for actual clinical scenarios.

[0049] Figure 2 Schematic diagram of the training process of the disease risk prediction model provided in the embodiment of this application. Figure 2 As shown in Figure 2, the training process of the disease risk prediction model is as follows:

[0050] Step 201: Obtain training sample data corresponding to the disease risk prediction model and supervision information corresponding to the training sample data.

[0051] Step 202: Obtain multiple teacher models pre-trained using training sample data and supervision information.

[0052] Step 203: Using the disease risk prediction model as the student model, the training sample data is input into multiple teacher models and the disease risk prediction model respectively to obtain the first probability distribution prediction results output by the multiple teacher models and the second probability distribution prediction results output by the disease risk prediction model.

[0053] Step 204: Determine a target loss function based on the first probability distribution prediction result and the second probability distribution prediction result.

[0054] Step 204 specifically includes: determining a cross-entropy loss function based on the first probability distribution prediction result and the second probability distribution prediction result; determining a time-dependent loss function based on the first probability distribution prediction result, the second probability distribution prediction result, a weight adjustment factor related to the time step, and a smoothness regularization term; and determining a target loss function based on the cross-entropy loss function and the time-dependent loss function. Determining the target loss function based on the cross-entropy loss function and the time-dependent loss function includes performing a weighted summation of the cross-entropy loss function and the time-dependent loss function to obtain the target loss function.

[0055] Step 205: Adjust the parameters of the disease risk prediction model according to the loss function.

[0056] Among them, the teacher model and the student model are machine learning models that integrate feature engineering. Feature engineering is used to perform feature selection, feature extraction and feature construction on the input of the machine learning model. Specifically, the features obtained based on feature engineering include: time series features, statistical features and change trend features; time series features are used to reflect the correlation between various physical examination indicators in historical physical examination information under time series; statistical features are used to reflect the overall distribution and discreteness of various physical examination indicators in historical physical examination information; change trend features are used to reflect the change trend of various physical examination indicators in historical physical examination information within a set time period. Among them, physical examination indicators refer to the above-mentioned blood pressure, cholesterol level, blood sugar, etc.

[0057] The following is a detailed description of the training process of the disease risk prediction model in stages:

[0058] Phase 1: Data Preprocessing

[0059] Obtain multiple training sample data, that is, the physical examination information of multiple users within a set time period, and preprocess these physical examination information. Specific preprocessing operations may be removing outliers, filling missing values, normalization, etc., to improve the data quality of the training sample data.

[0060] Phase 2: Feature Engineering

[0061] 1. Time series characteristics: Based on the time series of physical examination information of multiple users within a set time period, the correlation between the physical examination indicators of the historical physical examination information within the time series is determined. In specific implementation, the autocorrelation function (ACF) and partial autocorrelation coefficient (PACF) can be used to measure the correlation and periodicity of the time series:

[0062]

[0063]

[0064] Where τ represents the time lag, x t is the observed value of the physical examination index at time t, μ is the mean, represents the autocorrelation coefficient at lag τ.

[0065] It should be understood that ACF is a function used to calculate the correlation of a time series. It examines the correlation between the data in the time series and itself at different time points. It can be used to understand the strength of the correlation between different time points in the time series, and then determine how many past data points should be considered in the disease risk prediction model. PACF, on the other hand, is the correlation between the data at the current time point and the data at a certain time point in the past, given the data at other time points in the time series. It eliminates the influence of data at other time points and only focuses on the direct relationship between the current and past time points. Through ACF and PACF, we can understand the changes in various physical examination indicators in historical physical examination information over time, understand the correlation patterns between data in the time series, and select an appropriate disease risk prediction model accordingly.

[0066] Specifically, for example, if the ACF value is high, a complex time series model (such as ARIMA or SARIMA) can be used, while if the ACF value is low, a simple smoothing model or a moving average model can be used. Regarding PACF: a larger PACF value indicates a stronger direct correlation between the lags, and these lags should be considered in the model. A smaller PACF value indicates a weaker direct correlation between the lags, and these lags can be ignored.

[0067] 2. Statistical features: Descriptive statistical features, such as mean, median, standard deviation, and interquartile range, are extracted from the physical examination information of multiple users within a set time period. The specific formula is as follows:

[0068] μ=(1 / n)*∑(x t , t = 1 ton)

[0069]

[0070] Q1=x(n / 4), Q3=x(3n / 4)

[0071] IQR=Q3-Q1

[0072] Where n represents the number of physical examination index observations, x t is the observed value of the physical examination index at time t, μ represents the mean, σ represents the standard deviation, Q1 and Q3 represent the first and third quartiles respectively, and IQR represents the interquartile range.

[0073] It should be understood that the above statistical characteristics can reflect the overall distribution and dispersion of each physical examination indicator in the historical physical examination information.

[0074] 3. Change trend characteristics: Obtain the change trend of each physical examination indicator over time in the physical examination information of multiple users within a set time period, such as first-order difference, second-order difference, local extreme points, etc., to capture the dynamic changes of each physical examination indicator within the set time period. Specifically, the change speed and change acceleration can be used to measure the slope and curvature:

[0075] Δx t =x9t+1)-x t (First-order difference, indicating the rate of change)

[0076] Δ 2 x t =Δx(t+1)-Δx t (Second-order difference, indicating changing acceleration)

[0077] Among them, x t is the observed value of the physical examination index at time t. The first-order difference Δx-(t) measures the change rate of the physical examination index between time t and t+1, which reflects the short-term change trend of the physical examination index, such as the change of blood pressure within 3 days. The second-order difference Δ 2 x t It measures the speed of change of the change rate, that is, the acceleration of change, which reflects the dynamic adjustment of the change trend of physical examination indicators.

[0078] By comprehensively considering the change speed and acceleration of the above-mentioned physical examination indicators within the set time period, the static level and dynamic changes of each physical examination indicator are comprehensively portrayed, providing rich reference information for subsequent disease risk prediction, and thus laying the foundation for the training of disease risk prediction.

[0079] Phase 3: Building a Teacher Model

[0080] In an embodiment of the present application, in addition to the physical examination information of multiple users within a set time period, the training sample data may also include other clinical auxiliary information of multiple users, such as genetic data, medical history data, imaging data, etc. By using the above multimodal training sample data and corresponding supervision information to train the machine learning model of integrated feature engineering, a more accurate and stable teacher model can be obtained. Assuming that the training sample data is a historical record of various physical examination indicators of the user, the supervision information can be whether the user has experienced an event related to the target disease (such as brain disease).

[0081] In specific implementation, any of the following algorithms can be used for model training:

[0082] 1. Multiple Kernel Learning (MKL): This multi-kernel learning method can map data of different modalities to high-dimensional space through kernel functions for fusion. Specifically, assuming there are m datasets of different modalities, for the i-th modality, the corresponding kernel function K i Map the data to a high-dimensional space. Then, use the weight coefficient w i Perform linear combinations of the kernel matrices:

[0083]

[0084] Among them, K is the final fused kernel matrix, w i is the weight coefficient of the i-th kernel function, satisfying and w i ≥0. is the optimization weight coefficient w i , optimization algorithms such as stochastic gradient descent (SGD) or gradient boosted tree (GBT) can be used to achieve the best classification performance.

[0085] 2. Multi-Task Learning (MTL): The multi-task learning method can learn multiple related tasks at the same time and improve the generalization performance of the model by sharing information between tasks. In the embodiment of this application, it is assumed that there are n related tasks (for example, predicting the risk of different diseases). By introducing the weight matrix W∈R d×n , which can realize information sharing between tasks, where d is the feature dimension. For task i, it can be expressed as: Among them, y i is the output of the i-th task, X i is the input data of the i-th task, W i is the i-th column of the weight matrix $W$, is the noise term. To optimize the weight matrix W, regularization methods such as LASSO and Elastic Net can be used to minimize the loss function of all tasks.

[0086] 3. Deep Fusion Network (DFN): This method uses deep neural networks to extract and fuse features of multimodal data, which can achieve better expression ability and prediction performance.

[0087] The above are all specific examples and are not limited to these. Any machine learning or deep learning algorithm applicable to multimodal data can be used.

[0088] Stage 4: Knowledge Distillation

[0089] It should be understood that the purpose of knowledge distillation is to transfer the knowledge of a teacher model (usually a complex multimodal deep neural network) to a simpler student model (such as a shallow neural network or decision tree). In this way, the student model inherits the predictive power of the teacher model while reducing computing resource requirements, improving interpretability, and making it easier to deploy applications in real-world scenarios.

[0090] In specific implementations, the teacher model is trained using training sample data (at least including the physical examination information of the aforementioned multiple users within a set time period, as well as other clinical auxiliary information of the multiple users) to enable it to accurately predict the risk of disease. The training sample data is then input into multiple teacher models and a disease risk prediction model, respectively, to obtain first probability distribution prediction results output by the multiple teacher models and second probability distribution prediction results output by the disease risk prediction model. The first probability distribution prediction result is the probability distribution prediction result obtained by the multiple teacher models performing soft label prediction on the training sample data.

[0091] Based on the first probability distribution prediction result and the second probability distribution prediction result, determine the cross entropy loss function L CE (O (T) , O (S) ), and based on the first probability distribution prediction result, the second probability distribution prediction result, the weight adjustment factor related to the time step, and the smoothness regularization term, determine the time-related loss function L T (O (T) , O (S) ), the specific formula is as follows:

[0092]

[0093] in, Represents the teacher model output O of the i-th training sample at time step t (T) And the student model output O (S) The mean square error between t is the time weight parameter, f(t) is the weight adjustment factor, R(O (S) ) is the smoothness regularization term, and λ is the weight coefficient of the regularization term.

[0094] In related technologies, the time-correlation loss function generally uses a fixed time weight parameter, which fails to consider the dynamic impact of different time steps on the prediction results, resulting in a lower accuracy of the student model prediction. In view of this, this application makes the following improvements to the time-correlation loss function:

[0095] 1. A weight adjustment factor f(t) related to the time step t is introduced. It is used to dynamically adjust the weight parameters of each time step, and then adaptively adjust its contribution to the loss function according to the distance of the time step, better reflecting the relevance of the predictions of different time steps to the current moment. This adjustment factor f(t) can be an increasing or decreasing function with respect to t, or a function related to the rate of change of the data. For example, it can be defined as an exponential decay function:

[0096] f(t)=exp(-α·t)

[0097] Here, α is a hyperparameter that controls the decay rate, which can be set to 1, 0.5, or 0.1 according to experience, and is preferably 0.5 in the embodiment of the present application.

[0098] 2. A smoothness regularization term R(O (S) ), which is used to encourage the student model to learn a smoother prediction sequence. The regularization term is defined as the sum of squared differences between the outputs of adjacent time steps, as follows:

[0099]

[0100] In practical applications, minimizing the regularization term R(O (S) ), which can make the predicted values ​​of adjacent time steps as close as possible, thereby obtaining a smoother and continuous prediction sequence and improving the stability and interpretability of disease risk prediction.

[0101] It should be understood that in the process of knowledge distillation, the training goal of the student model is to obtain the target loss function, that is, to minimize the time-dependent loss function L T and conventional loss functions (such as cross entropy loss L CE ) is the weighted sum of:

[0102] L=β·L T (O (T) , O (S) )+(1-β)·L CE (O (T) , O (S) )

[0103] Among them, β is a weight parameter used to balance the two loss functions (time-related loss function L T and cross entropy loss L CE ) in optimizing model parameters. By minimizing the above loss function, the student model can achieve more accurate and stable prediction results while learning the teacher model's ability to model time series features, achieving better knowledge distillation.

[0104] After that, the parameters of the disease risk prediction model can be adjusted based on the target loss function.

[0105] Based on the above, the embodiment of the present application introduces the weight adjustment factor f(t), which can dynamically adjust the weight parameters of each time step according to the distance of the time step and the intensity of the data change, so that the disease risk prediction model can capture the time series characteristics more flexibly, thereby improving the accuracy of disease risk prediction. By introducing the smoothness regularization term R(O (S) ), the student model can learn a smoother and more continuous prediction sequence, improving the stability and interpretability of the prediction. In summary, the embodiment of the present application can reduce the complexity of the disease risk prediction model while ensuring the prediction performance, making it more suitable for actual clinical scenarios.

[0106] Furthermore, in order to ensure the performance of the trained student model (i.e., the disease risk prediction model), the method provided in the embodiment of the present application further includes:

[0107] Evaluate the performance of the student model;

[0108] Based on the results of performance evaluation, the student model is optimized.

[0109] In practical applications, multiple evaluation indicators can be selected to evaluate the performance of the student model. For example, the multiple evaluation indicators may include: Accuracy, Recall, Precision and F1-Score. It should be understood that these evaluation indicators can reflect the predictive ability of the disease risk prediction model from different angles. Among them, the accuracy rate is used to measure the correctness of the overall prediction of the disease risk prediction model; the recall rate is used to measure the ability of the disease risk prediction model to identify actual positive samples; the precision rate is used to measure the proportion of truly positive samples among the samples predicted as positive by the disease risk prediction model; the F1 value is the harmonic average of the recall rate and the precision rate, which takes both into consideration.

[0110] It's important to note that in machine learning, especially classification problems, positive samples are those labeled as positive by the classifier. In binary classification problems, one class is considered positive, and the other is considered negative. True positives are samples that are actually positive and correctly classified by the classifier, while false positives are samples that are actually negative but are mistakenly classified as positive.

[0111] In specific implementation, the accuracy is the ratio of the number of samples correctly predicted by the disease risk prediction model to the total number of samples. The specific calculation formula is:

[0112]

[0113] Among them, TP is the number of true positive samples, TN is the number of true negative samples, FP is the number of false positive samples, and FN is the number of false negative samples.

[0114] Recall: Also known as sensitivity, it refers to the ratio of the number of samples correctly predicted as positive by the disease risk prediction model to the total number of actual positive samples. The specific calculation formula is:

[0115]

[0116] It should be noted that the higher the recall rate, the stronger the disease risk prediction model's ability to identify actual positive samples.

[0117] Precision: refers to the proportion of truly positive samples among the samples predicted as positive by the disease risk prediction model. The specific calculation formula is:

[0118]

[0119] It should be noted that the higher the precision, the greater the proportion of actual positive samples among the samples predicted as positive by the disease risk prediction model.

[0120] F1-Score: It is the harmonic mean of recall and precision, taking both into account. Its specific calculation formula is:

[0121]

[0122] It should be noted that the higher the F1 value, the better the balance between recall and precision achieved by the disease risk prediction model, and the better the overall prediction performance.

[0123] In addition, evaluation metrics can also include ROC curves and AUC values. The ROC curve (Receiver Operating Characteristic Curve) is a graphical method for evaluating the performance of a binary classification model. It plots the changes in the true positive rate (TPR) and false positive rate (FPR) under different threshold settings to intuitively display the overall performance of the model. The AUC value (Area Under the Curve) is the area under the ROC curve, and its value range is [0,1]. The larger the AUC value, the better the performance of the disease risk prediction model.

[0124] Furthermore, the embodiment of the present application can use a validation set independent of the training set and the test set to evaluate the student model. Specifically, the validation set data is input into the student model to obtain a prediction result. Afterwards, the prediction performance and generalization ability of the student model on data that did not participate in the training are measured by calculating the above-mentioned evaluation indicators. By analyzing the evaluation indicators, the advantages and disadvantages of the model can be discovered, providing a basis for subsequent optimization. It should be noted that the training set, the test set, and the validation set can all include a user's historical physical examination information, but the data in the validation set is independent of the training set and the test set.

[0125] According to the above evaluation results, the student model can be optimized, and the optimization methods include but are not limited to adjusting the model structure, adjusting hyperparameters, optimizing the selection and regularization strategy, etc. When adjusting the model structure, targeted adjustments can be made based on the evaluation results of the above evaluation indicators. For example, when the accuracy is low, the fitting ability of the student model can be improved by increasing the number of hidden layers or hidden units. At the same time, regularization technology can be introduced to control the complexity of the student model to avoid overfitting. When the recall rate is too low, the weight of the positive samples can be increased by adjusting the category weights during the training process, so that the student model pays more attention to this part of the samples, and the number and diversity of positive samples can be increased through data enhancement, thereby improving the recall rate of the student model.

[0126] Afterwards, the optimized student model can be re-evaluated and optimized based on the re-evaluation results. This process can be repeated iteratively until the student model reaches the expected performance indicators.

[0127] After optimizing the student model, you can use the test set to evaluate it to test its generalization ability on unseen data. If the evaluation index on the test set is still high, it means that the student model has good disease risk prediction ability.

[0128] In summary, the disease risk prediction method provided in the embodiments of the present application provides new ideas and technical paths for early risk assessment and early warning of brain diseases, which is of great significance for reducing the incidence of brain diseases and alleviating the medical burden.

[0129] Figure 3 A schematic diagram of the structure of a disease risk prediction device provided in an embodiment of the present application is shown in FIG. Figure 3 As shown, the device includes: an acquisition module 31 and a prediction module 32.

[0130] An acquisition module 31 is used to acquire historical physical examination information of a target subject;

[0131] The prediction module 32 is used to input the historical physical examination information of the target object into the disease risk prediction model to obtain the disease risk prediction result corresponding to the target object; wherein, the disease risk prediction model is trained based on the time-related loss function, and the time-related loss function is determined based on the weight adjustment factor related to the time step and the smoothness regularization term.

[0132] Wherein, optionally, the device further includes: a training module, configured to obtain training sample data corresponding to the disease risk prediction model and supervision information corresponding to the training sample data;

[0133] Acquire multiple teacher models pre-trained using the training sample data; use the disease risk prediction model as a student model, input the training sample data into the multiple teacher models and the disease risk prediction model respectively, and obtain the first probability distribution prediction results output by the multiple teacher models and the second probability distribution prediction results output by the disease risk prediction model; determine the target loss function based on the first probability distribution prediction result and the second probability distribution prediction result; adjust the parameters of the disease risk prediction model according to the loss function.

[0134] Optionally, the teacher model and the student model are machine learning models that integrate feature engineering, and the feature engineering is used to perform feature selection, feature extraction and feature construction on the input of the machine learning model.

[0135] Among them, optionally, the features obtained based on the feature engineering include: time series features, statistical features and change trend features; the time series features are used to reflect the correlation between the physical examination indicators of the historical physical examination information under the time series; the statistical features are used to reflect the overall distribution and discreteness of each physical examination indicator in the historical physical examination information; the change trend features are used to reflect the change trend of each physical examination indicator in the historical physical examination information within the set time period.

[0136] Among them, optionally, the training module is also specifically used to: determine the cross entropy loss function based on the first probability distribution prediction result and the second probability distribution prediction result; determine the time-related loss function based on the first probability distribution prediction result, the second probability distribution prediction result, the weight adjustment factor related to the time step, and the smoothness regularization term; determine the target loss function based on the cross entropy loss function and the time-related loss function.

[0137] Among them, optionally, the training module is further used to: perform weighted summation on the cross entropy loss function and the time-related loss function to obtain a target loss function.

[0138] Optionally, the device further includes: an optimization module for performing performance evaluation on the student model; and optimizing the student model based on a result of the performance evaluation.

[0139] Figure 3 The device shown can execute the steps of the disease risk prediction method in the aforementioned embodiment. The detailed execution process and technical effects can be found in the description of the aforementioned embodiment and will not be repeated here.

[0140] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application is shown in FIG. Figure 4 As shown, the electronic device may include: a processor 41, a memory 42, and a communication interface 43. The memory 42 stores executable code, and when the executable code is executed by the processor 41, the processor 41 executes the disease risk prediction method in the above embodiment.

[0141] In addition, an embodiment of the present application provides a non-temporary machine-readable storage medium, on which executable code is stored. When the executable code is executed by a processor of an electronic device, the processor executes the disease risk prediction method in the aforementioned embodiment.

[0142] An embodiment of the present application provides a computer program product, which includes: a computer program, which, when executed by a processor of an electronic device, causes the processor to execute the disease risk prediction method in the aforementioned embodiment.

[0143] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0144] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0145] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0147] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0148] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0149] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0150] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0151] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A disease risk prediction method, characterized in that: include: Obtain historical physical examination information of the target subject; Inputting the historical physical examination information of the target subject into the disease risk prediction model to obtain the disease risk prediction result corresponding to the target subject; Among them, the disease risk prediction model is trained based on a time-related loss function, and the time-related loss function is determined based on a weight adjustment factor related to the time step and a smoothness regularization term.

2. The method according to claim 1, characterized in that The training process of the disease risk prediction model is as follows: Acquire training sample data corresponding to the disease risk prediction model and supervision information corresponding to the training sample data; Obtaining a plurality of teacher models pre-trained using the training sample data and the supervision information; Taking the disease risk prediction model as a student model, inputting the training sample data into the multiple teacher models and the disease risk prediction model respectively, obtaining first probability distribution prediction results output by the multiple teacher models and second probability distribution prediction results output by the disease risk prediction model; Determining a target loss function based on the first probability distribution prediction result and the second probability distribution prediction result; The parameters of the disease risk prediction model are adjusted according to the loss function.

3. The method according to claim 2, characterized in that The teacher model and the student model are machine learning models that integrate feature engineering, and the feature engineering is used to perform feature selection, feature extraction and feature construction on the input of the machine learning model.

4. The method according to claim 3, characterized in that The features obtained based on the feature engineering include: time series features, statistical features and change trend features; The time series feature is used to reflect the correlation between the various physical examination indicators of the historical physical examination information in the time series; The statistical features are used to reflect the overall distribution and dispersion of each physical examination indicator in the historical physical examination information; The change trend feature is used to reflect the change trend of each physical examination indicator in the historical physical examination information within a set time period.

5. The method according to claim 2, characterized in that The determining of a target loss function based on the first probability distribution prediction result and the second probability distribution prediction result includes: Determining a cross entropy loss function based on the first probability distribution prediction result and the second probability distribution prediction result; Determining the time-dependent loss function based on the first probability distribution prediction result, the second probability distribution prediction result, the weight adjustment factor related to the time step, and the smoothness regularization term; A target loss function is determined according to the cross entropy loss function and the time-dependent loss function.

6. The method according to claim 5, characterized in that The determining of a target loss function according to the cross entropy loss function and the time-dependent loss function includes: A weighted sum is performed on the cross entropy loss function and the time-dependent loss function to obtain a target loss function.

7. The method according to claim 2, characterized in that Also includes: performing a performance evaluation on the student model; Based on the results of the performance evaluation, the student model is optimized.

8. A disease risk prediction device, characterized in that: include: An acquisition module is used to obtain the historical physical examination information of the target object; A prediction module is used to input the historical physical examination information of the target subject into the disease risk prediction model to obtain the disease risk prediction result corresponding to the target subject; Among them, the disease risk prediction model is trained based on a time-related loss function, and the time-related loss function is determined based on a weight adjustment factor related to the time step and a smoothness regularization term.

9. An electronic device, characterized in that: include: A memory, a processor, and a communication interface; wherein the memory stores executable code, and when the executable code is executed by the processor, the processor executes the method according to any one of claims 1 to 7.

10. A non-transitory machine-readable storage medium, characterized in that The non-transitory machine-readable storage medium stores executable code, and when the executable code is executed by a processor of an electronic device, the processor is caused to perform the method according to any one of claims 1 to 7.

11. A computer program product, characterized in that include: A computer program, when executed by a processor of an electronic device, causes the processor to perform the method according to any one of claims 1 to 7.

Images