A neurology cognitive impairment prediction method and system
Patent Information
- Application Number
- CN202610969151.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-01
- Publication Date
- 2026-09-15
Smart Images
Figure CN122762285A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data analysis and prediction technology for cognitive impairment, specifically to, but not limited to, a method and system for predicting cognitive impairment in neurology. Background Technology
[0002] In neurology, early prediction of cognitive impairments, including Alzheimer's disease and vascular cognitive impairment, is crucial for delaying and controlling these diseases. Current clinical practice typically employs Extreme Gradient Boosting (XGBoost) ensemble learning models to process multimodal data associated with cognitive impairments in neurology patients. This yields the risk probability of cognitive impairment, which is then used to predict the degree and level of the impairment.
[0003] However, in practical applications, due to the insufficient number of clinical cognitive impairment samples used to train XGBoost, the accuracy of the risk probability obtained through the above processing of XGBoost is insufficient, thus failing to meet the prediction needs of cognitive impairment in neurology patients. Summary of the Invention
[0004] Based on the above technical problems, this application provides a method and system for predicting cognitive impairment in neurology, which can meet the needs of predicting cognitive impairment in neurology patients.
[0005] The technical solution provided in this application is as follows: This application provides a method for predicting cognitive impairment in neurology, including: Collect cognitive state data; wherein, the cognitive state data includes multimodal data related to the cognitive impairment status of patients in the neurology department; The cognitive state data is encoded by a variational autoencoder to obtain a clinical latent vector; wherein the clinical latent vector includes a low-dimensional embedding feature vector of the cognitive state data. The cognitive state data is processed by a conditional variational autoencoder to obtain a clinical feature vector; wherein the clinical feature vector includes the pathological features of the cognitive impairment carried by the cognitive state data. The clinical feature vector is predicted based on the full set of latent features by the decoder of the neural process network to obtain an initial prediction result; wherein, the full set of latent features includes clinical features corresponding to a set of clinical samples with a data volume greater than a certain threshold. The initial prediction result is corrected based on the clinical latent vector to obtain the target prediction result for the cognitive state data.
[0006] This application also provides a neurological cognitive impairment prediction system, including: The acquisition module is used to acquire cognitive state data; wherein, the cognitive state data includes multimodal data related to the cognitive impairment state of the patients in the neurology department; The processing module is used to encode the cognitive state data using an encoder of a variational autoencoder to obtain a clinical latent vector; wherein the clinical latent vector includes a low-dimensional embedding feature vector of the cognitive state data. The processing module is further configured to process the cognitive state data through a conditional variational autoencoder to obtain a clinical feature vector; wherein the clinical feature vector includes the pathological features of the cognitive impairment carried by the cognitive state data; The prediction module is used to predict the clinical feature vector based on the full set of latent features through the decoder of the neural process network to obtain an initial prediction result; wherein, the full set of latent features includes clinical features corresponding to a set of clinical samples with a data volume greater than a quantity threshold; The prediction module is further configured to correct the initial prediction result based on the clinical latent vector to obtain a target prediction result for the cognitive state data.
[0007] The technical solution provided in this application has at least the following beneficial effects: In the neurological cognitive impairment prediction method provided in this application embodiment, after collecting cognitive state data, including multimodal data related to the cognitive impairment state of neurological patients, the cognitive state data is encoded by the encoder of a variational autoencoder to obtain clinical latent vectors. These clinical latent vectors include low-dimensional embedded feature vectors of the cognitive state data. Thus, leveraging the advantages of continuous interpolation of the latent space and strong semantics of the latent features in the encoder of the variational autoencoder, the confidence of the clinical latent vectors can be improved. Furthermore, the cognitive state data is processed by a conditional variational autoencoder to obtain clinical feature vectors, which include the pathological features of cognitive impairment carried by the cognitive state data. Thus, leveraging the advantages of the conditional variational autoencoder in conditional reasoning, data augmentation, state transition, and conditional completion, the comprehensiveness and completeness of the features in the clinical feature vectors can be improved. Based on this, the decoder of the neural process network is used to analyze the full latent features... The method involves predicting clinical feature vectors to obtain initial prediction results. The full set of latent features includes clinical features corresponding to a clinical sample set with a data volume exceeding a threshold. Leveraging the stable probability prediction advantage of the neural process network decoder in dynamic scenarios and with a limited number of reference features, high-precision prediction of clinical feature vectors can still be achieved even when the number of clinical sample sets and their corresponding full set of latent features is insufficient, thus improving the accuracy of clinical feature vectors. Furthermore, the initial prediction results are corrected based on the clinical latent vectors to obtain target prediction results for cognitive state data, enabling targeted correction of the initial prediction results and improving the accuracy of the target prediction results. In summary, the technical solution provided in this application can improve the prediction accuracy of cognitive state data even when the number of clinical sample sets and full set of latent features is insufficient, thereby meeting the prediction needs for cognitive impairment in neurological patients. Attached Figure Description
[0008] Figure 1 A flowchart illustrating the neurological cognitive impairment prediction method provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the neurological cognitive impairment prediction system provided in the embodiments of this application. Detailed Implementation
[0009] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.
[0010] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.
[0011] In neurology, early prediction of cognitive impairments, including Alzheimer's disease and vascular cognitive impairment, is crucial for delaying and controlling these diseases. Current clinical practice typically employs XGBoost ensemble learning models to process multimodal data associated with cognitive impairments in neurology patients, obtaining the risk probability of each patient's cognitive impairment, and then predicting the degree and level of the impairment based on this risk probability.
[0012] However, in practical applications, due to the insufficient number of clinical cognitive impairment samples used to train XGBoost, the accuracy of the risk probability obtained through the above processing of XGBoost is insufficient, thus failing to meet the prediction needs of cognitive impairment in neurology patients.
[0013] Based on the above problems, this application provides a method and system for predicting cognitive impairment in neurology. Figure 1 This is a flowchart illustrating the neurological cognitive impairment prediction method provided in the embodiments of this application, as shown below. Figure 1 As shown, the method may include the following steps: Step 101: Collect cognitive state data.
[0014] The cognitive state data includes multimodal data related to the cognitive impairment status of patients in the neurology department.
[0015] In some embodiments, cognitive impairment states may include mild cognitive impairment, acute cognitive impairment, and cognitive impairment of specific etiologies.
[0016] In some embodiments, multimodal data may include at least two types of data associated with cognitive impairment states; exemplaryly, multimodal data may include clinical rating scales, medical images, and biomarkers; exemplaryly, medical images may include magnetic resonance gray matter volume and fluorodeoxyglucose-positron emission tomography-normalized uptake values for patients; exemplaryly, biomarkers may include blood markers, which may include phosphorylated tau protein 181, β-amyloid 42 / 40 ratio, and neurofilament light chain protein; exemplaryly, clinical rating scales may include the Mini-Mental State Examination, the Montreal Cognitive Assessment Scale, and the Activities of Daily Living Scale.
[0017] In some embodiments, multimodal data may also include basic patient information; for example, basic information may include the patient's age, gender, apolipoprotein E genotype, dietary and lifestyle habits, and family history of genetic diseases.
[0018] Step 102: Encode the cognitive state data using the encoder of the variational autoencoder to obtain the clinical latent vector.
[0019] Among them, the clinical latent vectors include low-dimensional embedding feature vectors of cognitive state data.
[0020] In some embodiments, before encoding the cognitive state data using the encoder of a variational autoencoder, the cognitive state data can be preprocessed to obtain preprocessed data, and then encoded using the encoder of the variational autoencoder to obtain clinical latent vectors.
[0021] For example, in actual clinical applications, different testing organizations may have different equipment types, data acquisition methods, testing methods, and data storage methods for collecting cognitive status data. Therefore, after obtaining cognitive status data, the source identifier associated with the cognitive status data can be obtained simultaneously, and preprocessing can be performed on the cognitive status data based on the source identifier. For example, the source identifier may include the number or identifier of the testing organization that provides the cognitive status data. For example, the testing organization may include the laboratory department of a hospital or a medical testing institution.
[0022] Specifically, preprocessing can be achieved in the following ways: Obtain the preprocessing baseline data corresponding to the source identifier, and then perform preprocessing on the cognitive state data based on the preprocessing baseline data; for example, the preprocessing baseline data may include mean data and variance data, wherein the mean data may include the mean obtained by statistically averaging the set of cognitive state data collected by the detection organization in multiple historical periods, and the variance data may include the variance obtained by statistically averaging the set of cognitive state data collected by the detection organization in multiple historical periods; specifically as shown in equation (1): (1) in, For the preprocessed data and the d-th dimension of the i-th data in the multimodal data The corresponding data, For the m-th source identifier Related and with The corresponding mean data, To and The corresponding variance data, Let m be the amount of data in the cognitive state dataset associated with the m-th detection organization corresponding to the k-th source identifier. It is a very small number, used to reduce the probability of division by zero exceptions. For the d-th dimension of the m-th source identifier, For the j-th source identifier, The d-th dimension data is used to identify the j-th source; the i-th data can be associated with the i-th patient; m, i, and j are all integers greater than or equal to 1.
[0023] In some embodiments, the clinical latent vector may include a set of cognitive impairment features carried by cognitive state data.
[0024] In some embodiments, the clinical latent vector may include a first distribution state consisting of a first latent mean and a first latent variance; specifically, the first latent mean and the first latent variance may constitute a probability distribution of a low-dimensional latent space represented by cognitive state data, so as to reflect the core cognitive impairment features carried by the cognitive state data; wherein, the first latent mean may represent the uncertainty of the cognitive state data, and the first latent variance may represent the fluctuation range of the cognitive state data.
[0025] For example, the first distribution state may include the essential latent representation represented by the cognitive state data; correspondingly, the decoder of the variational autoencoder can reconstruct the cognitive state data based on the first distribution state.
[0026] For example, cognitive state data can be input into the encoder of a variational autoencoder to encode the cognitive state data to obtain a clinical latent vector; for example, the clinical latent vector may also include a popular embedding obtained by sampling a first distribution state expressed by a first latent mean and a first latent variance.
[0027] Step 103: Process the cognitive state data using a conditional variational autoencoder to obtain a clinical feature vector.
[0028] The clinical feature vector includes the pathological features of cognitive impairment carried by the cognitive state data.
[0029] In some embodiments, the number of features contained in the clinical feature vector may be greater than the number of features contained in the clinical latent vector, or the dimension of the features contained in the clinical feature vector may be greater than the dimension of the features contained in the clinical latent vector.
[0030] In some embodiments, the clinical feature vector can be obtained in the following way: Conditional data corresponding to cognitive state data is constructed. Then, the conditional data and cognitive state data are processed by the encoder of the conditional variational autoencoder to obtain the second distribution state. The second distribution state is then decoded and reconstructed based on the conditional data by the decoder of the conditional variational autoencoder to obtain the clinical feature vector.
[0031] For example, the second distribution state can be described by the second latent mean and the second latent variance, which can characterize the posterior distribution of the latent variables corresponding to the cognitive state data under the conditions described by the conditional data.
[0032] For example, the conditional data may include at least one condition constructed to extend the cognitive state data; for example, the conditional data may be determined by the gap between the cognitive state data and reference state data, which may include the full range of state data required for a clinical diagnosis of cognitive impairment.
[0033] Step 104: The clinical feature vector is predicted based on the full set of latent features by the decoder of the neural process network to obtain the initial prediction result.
[0034] Among them, the full potential features include clinical features corresponding to clinical sample sets with a data volume greater than the quantity threshold.
[0035] In some embodiments, the clinical sample set may include a set of sample data on the cognitive impairment status of neurological patients collected in advance over multiple historical periods; for example, the integrity of the data in the clinical sample set may be greater than the integrity threshold, that is, the number of modalities carried by the sample data in the clinical sample set may be greater than the first threshold, and the feature dimension of the data of each modality may be greater than or equal to the second threshold.
[0036] In some embodiments, the full set of potential features may include a set of features carried by sample data in a clinical sample set that are used to comprehensively characterize cognitive impairment.
[0037] In some embodiments, the full set of latent characteristics can be obtained in advance by analyzing a clinical sample set; specifically, the full set of latent characteristics can be obtained in the following ways: The full set of latent features is obtained by analyzing sample data from a clinical sample set through an encoder of a neural process network.
[0038] In some embodiments, the initial prediction results may include a preliminarily determined degree, risk, or level of cognitive impairment corresponding to cognitive state data, and may also include a probability corresponding to the aforementioned degree, risk, or level.
[0039] In some embodiments, the initial prediction results can be obtained in the following ways: The clinical feature vector and all potential features are decoded and mapped by the decoder of the neural process network to obtain the third distribution state, and the initial prediction result is determined based on the third distribution state; wherein the prediction mean can be greater than or equal to 0 and less than or equal to 1.
[0040] Specifically, the third distribution state can include a Gaussian distribution state consisting of the predicted mean and the predicted variance; correspondingly, the initial prediction results can be obtained in the following way: The initial prediction result is obtained by performing the calculation of the predicted mean using the logit function as shown in equation (2): (2) in, To predict the mean, This is the initial prediction result.
[0041] Step 105: Correct the initial prediction results based on the clinical latent vector to obtain the target prediction results for the cognitive state data.
[0042] In some embodiments, the target prediction result may be the same as or different from the initial prediction result; for example, the target prediction result may characterize the probability of at least one degree, risk or level of cognitive impairment corresponding to the cognitive state data in the form of a probability.
[0043] In some embodiments, the target prediction result can be obtained in the following way: The decoder of the neural process network predicts the clinical latent vector based on the full set of latent features to obtain a corrected prediction result. If the difference between the corrected prediction result and the initial prediction result is greater than or equal to the difference threshold, the corrected prediction result and the initial prediction result are statistically averaged, and the result of the statistical average is determined as the target prediction result. If the difference between the corrected prediction result and the initial prediction result is less than the difference threshold, a correction coefficient can be determined based on the difference between the corrected prediction result and the initial prediction result can be corrected based on the correction coefficient, and the corrected result is determined as the target prediction result. If the corrected prediction result is consistent with the initial prediction result, the initial prediction result can be determined as the target prediction result.
[0044] As can be seen from the above, in the neurological cognitive impairment prediction method provided in this application embodiment, after collecting cognitive state data, including multimodal data related to the cognitive impairment state of neurological patients, the cognitive state data is encoded by the encoder of a variational autoencoder to obtain clinical latent vectors. These clinical latent vectors include low-dimensional embedded feature vectors of the cognitive state data. Thus, by leveraging the continuous interpolation of the latent space and the strong semantics of the latent features in the encoder of the variational autoencoder, the confidence of the clinical latent vectors can be improved. Furthermore, by processing the cognitive state data through a conditional variational autoencoder, clinical feature vectors are obtained. These clinical feature vectors include the pathological features of cognitive impairment carried by the cognitive state data. Thus, by leveraging the advantages of the conditional variational autoencoder in conditional reasoning, data augmentation, state transition, and conditional completion, the comprehensiveness and completeness of the features in the clinical feature vectors can be improved. Based on this, the decoder of the neural process network is used to analyze the full dataset. Latent features are used to predict clinical feature vectors to obtain initial prediction results. The full set of latent features includes clinical features corresponding to a clinical sample set with a data volume exceeding a threshold. Thus, leveraging the stable probability prediction advantage of the neural process network decoder in dynamic scenarios and with a limited number of reference features, high-precision prediction of clinical feature vectors can still be achieved even when the number of clinical sample sets and their corresponding full set of latent features is insufficient, thereby improving the accuracy of clinical feature vectors. Furthermore, the initial prediction results are corrected based on the clinical latent vectors to obtain target prediction results for cognitive state data, enabling targeted correction of the initial prediction results and improving the accuracy of the target prediction results. In summary, the technical solution provided in this application can improve the prediction accuracy of cognitive state data even when the number of clinical sample sets and full set of latent features is insufficient, thereby meeting the prediction needs for cognitive impairment in neurological patients.
[0045] Based on the foregoing embodiments, the neurological cognitive impairment prediction method provided in this application, which corrects the initial prediction result based on the clinical latent vector to obtain the target prediction result for cognitive state data, can be achieved through the following steps: Step A1: Fuse the clinical latent vector and the clinical feature vector to obtain the clinical fusion result.
[0046] In some embodiments, clinical fusion results can be obtained in the following ways: The clinical latent vectors are mapped using a lightweight first encoder to obtain first data, and the clinical feature vectors are mapped using a lightweight second encoder to obtain second data. The first data and the second data are then concatenated and fused to obtain the clinical fusion result. For example, the first encoder may include at least one layer of multi-layer perceptron (MLP), and the second encoder may include at least two layers of MLP.
[0047] Specifically, the first data can be obtained in the following way: The samples in the clinical sample set are pre-processed using a variational autoencoder to obtain a sample latent vector set. Then, the Euclidean distance between the clinical latent vectors and the latent vectors in the sample latent vector set is calculated using the k-nearest neighbor counting method to obtain a Euclidean distance set. Next, the average distance of the first k distances in the Euclidean distance set is taken, and based on the average distance value, the vectors in the clinical latent vector set are divided into a dense vector set and a sparse vector set. The dense vector set is mapped using a lightweight first encoder to obtain the mapping result, and the mapping result and the sparse vector set are determined as the first data. The value of k can be preset or adjusted, and this embodiment does not limit it.
[0048] Step A2: Process the clinical fusion results using a temperature network to obtain the temperature scaling factor.
[0049] In some embodiments, temperature scaling factor It can be calculated using equation (3): (3) in, As the first data, This is the second data point.
[0050] Step A3: Process the initial prediction results based on the temperature scaling factor to obtain the target prediction results.
[0051] In some embodiments, the target prediction result It can be calculated using equation (4): (4) in, This is the sigmoid function.
[0052] As can be seen from the above, the neurological cognitive impairment prediction method provided in this application obtains a clinical fusion result by fusing clinical latent vectors and clinical feature vectors, thereby achieving a comprehensive fusion of clinical latent vectors and clinical feature vectors and improving the comprehensiveness of the clinical fusion result. Furthermore, by processing the clinical fusion result through a temperature network, a temperature scaling factor is obtained. In this way, by leveraging the advantage of the temperature network in adjusting the temperature scaling factor according to the clinical fusion result, the accuracy of the temperature scaling factor can be improved. On this basis, the initial prediction result is processed based on the temperature scaling factor to obtain the target prediction result, which enables targeted processing of the initial prediction result and thus improves the accuracy of the target prediction result.
[0053] Based on the foregoing embodiments, the neurological cognitive impairment prediction method provided in this application, which processes cognitive state data through a conditional variational autoencoder to obtain a clinical feature vector, can be achieved through the following steps: Step B1: Determine the clinical mask data corresponding to the cognitive state data.
[0054] In some embodiments, clinical mask data can characterize the quantity and type of missing data in cognitive state data; exemplarily, clinical mask data may include conditional data as described in the foregoing embodiments; accordingly, clinical mask data can be determined in the following ways: The cognitive state data and reference state data are analyzed to identify missing state data in the cognitive state data. Based on the arrangement of data in the reference state data, the data identifiers corresponding to the missing state data are set as missing (represented by the number 0), while the data identifiers corresponding to the cognitive state data can be set as non-missing (represented by the number 1). The set of the above data identifiers is determined as the clinical mask data.
[0055] Step B2: The conditional variational autoencoder encodes the cognitive state data based on the clinical mask data to obtain the latent variable distribution data.
[0056] In some embodiments, the latent variable distribution data may include the second distribution state in the foregoing embodiments, which can be characterized by the second latent mean and the second latent variance; specifically, the latent variable distribution data can be calculated by equation (5): (5) in, This is the second hidden mean. This is the second hidden variance. Here, c represents the descriptive data for the latent variable distribution, and c represents the clinical mask data. This is preprocessed data corresponding to cognitive state data. This refers to the encoder's computation function in a conditional variational autoencoder.
[0057] Step B3: The latent variable distribution data and cognitive state data are processed by the decoder of the conditional variational autoencoder to obtain the clinical feature vector.
[0058] For example, the clinical feature vector can be calculated using equation (6): (6) in, For clinical feature vectors, For the latent variable distribution data corresponding to the second distribution state, This is the operation function of the decoder for a conditional variational autoencoder.
[0059] As can be seen from the above, in the neurological cognitive impairment prediction method provided in this application embodiment, clinical mask data corresponding to cognitive state data is determined, and the encoder of the conditional variational autoencoder encodes the cognitive state data based on the clinical mask data to obtain latent variable distribution data. In this way, targeted encoding processing of cognitive state data can be achieved, improving the correlation between latent variable distribution data, clinical mask data, and cognitive state data. Furthermore, the decoder of the conditional variational autoencoder processes the latent variable distribution data and cognitive state data to obtain clinical feature vectors, achieving targeted augmentation of cognitive state data, thereby improving the comprehensiveness and completeness of data in the clinical feature vectors.
[0060] Based on the foregoing embodiments, the method for predicting cognitive impairment in neurology provided in this application can determine the clinical mask data corresponding to the cognitive state data through the following steps: Step C1: Obtain the source identifier associated with the cognitive state data.
[0061] In some embodiments, the source identifier can be an attribute tag of the cognitive state data, and the source identifier can be associated with the cognitive state data. Therefore, after obtaining the cognitive state data, the source identifier associated with the cognitive state data can be obtained.
[0062] In some embodiments, the source identifier may characterize the name or number of the detection organization that collected the cognitive state data.
[0063] Step C2: Obtain the full target data corresponding to the source identifier.
[0064] In some embodiments, the full clinical data corresponding to different testing organizations may be different; for example, the full clinical data may include all the clinical data, features and indicator values required in clinical practice to diagnose and track cognitive impairment.
[0065] In one embodiment, the association between the source identifier in the identifier set and the clinical full data in the clinical full data set can be established in advance. In this way, after the source identifier is actually obtained, the clinical full data corresponding to the actual source identifier can be obtained from the clinical full data according to the degree of matching between the source identifier and the source identifier in the identifier set, and the clinical full data is determined as the target full data.
[0066] Step C3: Determine the clinical mask data based on the degree of difference between the target full data and the cognitive state data.
[0067] In some embodiments, clinical mask data can be determined in the following ways: The data identifiers corresponding to missing data included in the target full dataset but not included in the cognitive state data are set to 0, while the data identifiers corresponding to data not included in either the cognitive state data or the target full dataset are set to 1. The set of data identifiers is then determined as the clinical mask data.
[0068] As can be seen from the above, in the neurological cognitive impairment prediction method provided in this application embodiment, the source identifier corresponding to the cognitive state data is obtained, and the target full data corresponding to the source identifier is obtained. In this way, the association between the cognitive state data and the target full data is realized through the source identifier. On this basis, the clinical mask data is determined based on the degree of difference between the target full data and the cognitive state data, which can improve the comprehensiveness and accuracy of the clinical mask data.
[0069] Based on the foregoing embodiments, in the neurological cognitive impairment prediction method provided in this application, before the decoder of the neural process network predicts the clinical feature vector based on the full set of latent features, the following steps may also be performed: Step D1: Encode the clinical sample set using the encoder of the neural process network to obtain the sample feature set.
[0070] In some embodiments, the clinical sample set may include cognitive impairment labels; for example, the cognitive impairment labels may characterize the type, severity, and developmental status of the cognitive impairment represented by the sample data in the clinical sample set.
[0071] In some embodiments, the nth sample data in the sample feature set Corresponding nth sample features It can be calculated using equation (7): (7) in, The cognitive impairment label corresponding to the nth sample data. This is the computation function for the encoder of the neural process network.
[0072] Step D2: Integrate the sample feature set to obtain the full set of latent features.
[0073] For example, the full set of latent features can be obtained by statistically averaging the sample features in the sample feature set.
[0074] As can be seen from the above, in the neurological cognitive impairment prediction method provided in this application, the clinical sample set is encoded by the encoder of the neural process network to obtain the sample feature set, and the sample feature set is integrated to obtain the full set of latent features. In this way, by leveraging the advantage of the encoder of the neural process network in tracking and integrating scattered features, the comprehensiveness of the sample feature set can be improved, thereby improving the integrity of the full set of latent features.
[0075] Based on the foregoing embodiments, in the neurological cognitive impairment prediction method provided in this application, before encoding the clinical sample set through a neural process network to obtain the sample feature set, the following operations may also be performed: The clinical sample set is divided into a support set and a target set; the initial neural process network is trained based on the support set and the target set to obtain the neural process network.
[0076] In some embodiments, the number of sample data contained in the support set can be the same as the number of sample data contained in the target set. That is, the clinical sample set can be evenly divided to obtain the support set and the target set.
[0077] In some embodiments, the clinical sample set may include sample data and sample distribution labels; wherein, the sample distribution labels are used to characterize the sample distribution parameters corresponding to the sample data; specifically, the sample distribution parameters may include the sample mean and the sample variance.
[0078] For example, sample distribution labels can characterize the distribution state of sample data in a clinical sample set; for example, sample distribution labels can be pre-annotated by professionals or obtained by processing the clinical sample set through a conditional variational autoencoder.
[0079] For example, the parameters of the initial neural process network can be adjusted based on the sample data and sample distribution labels contained in the support set to obtain the parameter-adjusted initial neural process network.
[0080] For example, the process of encoding the support set by the encoder of the initial neural network can be the same as the process of encoding the clinical sample set by the encoder of the neural process network in the previous embodiment; correspondingly, the encoding process performed by the decoder of the initial neural process network can be the same as the process of decoding and mapping the clinical feature vector and the full set of latent features by the decoder of the neural process network in the previous embodiment, and will not be described again here.
[0081] For example, the sample data contained in the target set can be processed by an initial neural process network with adjusted parameters to obtain the predicted distribution parameters. Then, the prediction accuracy is optimized based on the negative log-likelihood between the predicted distribution parameters and the sample distribution parameters, while the uncertainty is calibrated with a regularization term.
[0082] For example, the parameter adjustment process of the initial neural network can be carried out using a meta-learning paradigm.
[0083] For example, the prediction accuracy is optimized based on the negative log-likelihood between the predicted distribution parameters and the sample distribution parameters, while the first loss corresponding to the uncertainty is calibrated with a regularization term. It can be characterized by equation (8): (8) Where P is the support set, Q is the target set, ss is the total latent features represented by the sample data to be predicted in the target set SS, and y is the binary classification label of the sample data to be predicted in the target set. The predicted probability is obtained from the sample data. This represents the uncertainty variance of the initial neural network output. and These constitute the predicted distribution parameters. These are the uncertainty regularization coefficients, used to control the regularization term. The weight, This is the expectation operator.
[0084] As can be seen from the above, in the neurological cognitive impairment prediction method provided in this application, the clinical sample set is divided into a support set and a target set. An initial neural process network is trained based on the support set and the target set to obtain the neural process network. Thus, by using the support set and the target set to train the initial neural process network, the neural process network obtained through the above training process possesses advantages such as few-sample learning, uncertainty calibration, permutation invariance, and strong generalization ability. This enables the neural process network to efficiently and accurately process complex and variable cognitive impairment-related state and feature data.
[0085] Based on the foregoing embodiments, in the neurological cognitive impairment prediction method provided in this application, before processing the cognitive state data through a conditional variational autoencoder to obtain the clinical feature vector, the following operations may also be performed: Obtain the sample mask data corresponding to the initial sample set; augment the initial sample set based on the sample mask data using a conditional variational autoencoder to obtain the clinical sample set.
[0086] In some embodiments, the conditional variational autoencoder can be obtained by pre-training an initial conditional variational autoencoder based on a target sample set.
[0087] In some embodiments, the target sample data may include a set of clinical cognitive state data with completeness greater than or equal to a completeness threshold; for example, the target sample data may be pre-collected.
[0088] In some embodiments, the sample mask data can be pre-constructed or randomly constructed. For example, the sample mask data is used to mask the target sample data in the target sample set to obtain the missing samples and observed samples corresponding to the target sample data. The observed samples can carry data or information of at least one dimension of cognitive impairment, but do not carry all the data in the target sample data. That is, the observed samples and the missing samples can both be proper subsets of the initial sample data, and the missing samples can be complementary to the observed samples. The union of the missing samples and the observed samples can constitute the target sample data.
[0089] Specifically, the training process for the initial conditional variational autoencoder can be as follows: First, the target sample data is divided into observed samples and missing samples based on the sample mask data. The encoder of the initial conditional variational autoencoder processes the observed samples and sample mask data to obtain the initial mean and initial variance of the latent variable distribution corresponding to the target sample data. Next, the decoder of the initial conditional variational autoencoder processes the aforementioned latent variables and sample mask data to generate predicted sample data. Finally, based on the second loss between the predicted sample data and the target sample data, the parameters of the initial conditional variational autoencoder are iteratively optimized and adjusted to obtain the conditional variational autoencoder. Specifically, the second loss... It can be calculated using the following formula: in, For the observation sample corresponding to the i-th target sample data, For the missing samples corresponding to the i-th target sample data, These are the initial latent variables of the encoder output of the initial conditional variational autoencoder. This refers to the encoder's computation function in the initial condition variational autoencoder. This is the computational function of the decoder in the initial conditional variational autoencoder. Z represents the prior distribution of the initial latent variables, typically a standard normal distribution, where Z is an integer greater than 1, representing the total number of initial sample data in the initial sample set. Here is the formula for calculating the KL divergence.
[0090] In some embodiments, the clinical sample set can be obtained in the following ways: Matching analysis is performed on the initial sample data and reference state data in the initial sample set to determine the target mask data corresponding to the initial sample data in the initial sample set. Then, the initial sample data is augmented based on the target mask data by a variational autoencoder to obtain the augmented data corresponding to the initial sample data. This process is repeated to traverse each initial sample data in the initial sample set to obtain the augmented data corresponding to each initial sample data. The set of augmented data is then determined as the clinical sample set.
[0091] As described above, the neurological cognitive impairment prediction method provided in this application obtains sample mask data corresponding to the initial sample set, and then uses a conditional variational autoencoder to augment the initial sample set based on the sample mask data to obtain a clinical sample set. Thus, through the above operations, targeted augmentation of the initial sample set can be achieved even when the initial sample set has some missing data, thereby improving the comprehensiveness and completeness of the data in the clinical sample set.
[0092] To verify the actual clinical efficacy of this solution, this application provides a comparison of the clinical treatment effects of a comparative solution and this solution. The comparative solution may include a first solution and a second solution, and the comparative data includes cognitive status data from 800 cases from different medical institutions. It should be noted that the above analysis and comparison of the cognitive status data of 800 cases were performed with the permission of the patients and medical institutions.
[0093] Specifically, the first approach uses a logistic regression prediction model to evaluate and predict cognitive state data; the second approach uses an XGBoost ensemble learning model to process cognitive state data.
[0094] The comparison results between this scheme and the comparative scheme are shown in Table 1:
[0095] Table 1 As shown in Table 1, under conditions of full comparative data, data corresponding to the mild cognitive impairment subgroup, data corresponding to the apolipoprotein E ε4 carrier subgroup, data corresponding to missing multimodal data, and intercenter migration, the accuracy of the predicted risk provided by the technical solution in this application is improved compared with comparative scheme 1 and comparative scheme 2. Among them, the data corresponding to the mild cognitive impairment subgroup includes part of the comparative data corresponding to the mild cognitive impairment subgroup, the data corresponding to the apolipoprotein E ε4 carrier subgroup includes part of the comparative data corresponding to the apolipoprotein E ε4 carrier subgroup, and the data corresponding to missing multimodal data includes comparative data with a multimodal loss rate greater than or equal to 20%.
[0096] Table 2 provides a comparison of the effectiveness of the proposed scheme with the proposed scheme across multiple evaluation metrics.
[0097]
[0098] Table 2 As can be seen from Table 2, this scheme improves upon both Comparison Scheme 1 and Comparison Scheme 2 in terms of calibration error, maximum calibration error, Brier score, and sparse region confidence. Among these, sparse region confidence can include cases where the predicted risk probability is less than the probability threshold but the actual risk is high.
[0099] Table 3 presents the statistical results of ablation experiments provided in the embodiments of this application.
[0100]
[0101] Table 3 In Table 3, the baseline architecture is the architecture of the comparison scheme 1. The first architecture is obtained by adding a variational autoencoder to the baseline architecture. The second architecture is obtained by adding a neural process network to the first architecture. The third architecture is obtained by adding a conditional variational autoencoder to the first architecture. The fourth architecture is obtained by adding a neural process network to the third architecture. The full architecture is the complete architecture of the technical solution provided in this application.
[0102] As shown in Table 3, the addition of at least one data processing step from variational autoencoders, neural process networks, and conditional variational autoencoders can improve dimensions such as the area under the curve (AUC), missing sample AUC, calibration error, number of parameters, and inference time. Therefore, the various data processing steps in the technical solution provided in this application, through organic integration, can improve the effectiveness of clinical cognitive impairment prediction.
[0103] Based on the foregoing embodiments, this application also provides a neurological cognitive impairment prediction system. Figure 2This is a schematic diagram of the structure of the neurological cognitive impairment prediction system provided in the embodiments of this application, such as... Figure 2 As shown, the neurological cognitive impairment prediction system 200 may include: The acquisition module 201 is used to acquire cognitive state data; wherein, the cognitive state data includes multimodal data associated with the cognitive impairment status of patients in the neurology department; Processing module 202 is used to encode the cognitive state data through the encoder of the variational autoencoder to obtain the clinical latent vector; wherein the clinical latent vector includes the low-dimensional embedding feature vector of the cognitive state data. The processing module 202 is also used to process the cognitive state data through a conditional variational autoencoder to obtain a clinical feature vector; wherein the clinical feature vector includes the pathological features of cognitive impairment carried by the cognitive state data; The prediction module 203 is used to predict the clinical feature vector based on the full set of latent features through the decoder of the neural process network to obtain the initial prediction result; wherein, the full set of latent features includes the clinical features corresponding to the clinical sample set with a data volume greater than the number threshold. The prediction module 203 is also used to correct the initial prediction results based on the clinical latent vector to obtain the target prediction results for the cognitive state data.
[0104] In some embodiments, the prediction module is used to fuse clinical latent vectors and clinical feature vectors to obtain a clinical fusion result; process the clinical fusion result through a temperature network to obtain a temperature scaling factor; and process the initial prediction result based on the temperature scaling factor to obtain a target prediction result.
[0105] In some embodiments, the processing module is configured to determine the clinical mask data corresponding to the cognitive state data; encode the cognitive state data based on the clinical mask data using the encoder of a conditional variational autoencoder to obtain latent variable distribution data; and process the latent variable distribution data and the cognitive state data using the decoder of the conditional variational autoencoder to obtain a clinical feature vector.
[0106] In some embodiments, the processing module is configured to obtain a source identifier associated with the cognitive state data; obtain the target full data corresponding to the source identifier; and determine clinical mask data based on the degree of difference between the target full data and the cognitive state data.
[0107] In some embodiments, the processing module is configured to encode a set of clinical samples using an encoder of a neural process network to obtain a set of sample features; The sample feature set is integrated and processed to obtain the full set of latent features.
[0108] In some embodiments, the processing module is configured to divide the clinical sample set into a support set and a target set; and to train an initial neural process network based on the support set and the target set to obtain a neural process network.
[0109] In some embodiments, the processing module is configured to acquire sample mask data corresponding to the initial sample set; and to perform augmentation processing on the initial sample set based on the sample mask data using a conditional variational autoencoder to obtain a clinical sample set.
[0110] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.
[0111] The methods disclosed in the various method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict.
[0112] The features disclosed in the various product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.
[0113] The features disclosed in the various method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.
[0114] It should be noted that the aforementioned computer-readable storage media can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various electronic devices that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.
[0115] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0116] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware nodes. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.
[0118] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0121] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A method for predicting cognitive impairment in neurology, characterized in that, The method includes: Collect cognitive state data; wherein, the cognitive state data includes multimodal data related to the cognitive impairment status of patients in the neurology department; The cognitive state data is encoded by a variational autoencoder to obtain a clinical latent vector; wherein the clinical latent vector includes a low-dimensional embedding feature vector of the cognitive state data. The cognitive state data is processed by a conditional variational autoencoder to obtain a clinical feature vector; wherein the clinical feature vector includes the pathological features of cognitive impairment carried by the cognitive state data; The clinical feature vector is predicted based on the full set of latent features by the decoder of the neural process network to obtain an initial prediction result; wherein, the full set of latent features includes clinical features corresponding to a set of clinical samples with a data volume greater than a certain threshold. The initial prediction result is corrected based on the clinical latent vector to obtain the target prediction result for the cognitive state data.
2. The method according to claim 1, characterized in that, The step of correcting the initial prediction result based on the clinical latent vector to obtain the target prediction result for the cognitive state data includes: The clinical latent vector and the clinical feature vector are fused to obtain the clinical fusion result; The clinical fusion results are processed using a temperature network to obtain a temperature scaling factor; The initial prediction result is processed based on the temperature scaling factor to obtain the target prediction result.
3. The method according to claim 1, characterized in that, The process of processing the cognitive state data using a conditional variational autoencoder to obtain a clinical feature vector includes: Determine the clinical mask data corresponding to the cognitive state data; The conditional variational autoencoder encodes the cognitive state data based on the clinical mask data to obtain latent variable distribution data. The latent variable distribution data and the cognitive state data are processed by the decoder of the conditional variational autoencoder to obtain the clinical feature vector.
4. The method according to claim 3, characterized in that, The step of determining the clinical mask data corresponding to the cognitive state data includes: Obtain the source identifier associated with the cognitive state data; Obtain the full target data corresponding to the source identifier; The clinical mask data is determined based on the degree of difference between the target full data and the cognitive state data.
5. The method according to claim 1, characterized in that, Before the decoder of the neural process network predicts the clinical feature vector based on the full set of latent features, the method further includes: The clinical sample set is encoded by the encoder of the neural process network to obtain a sample feature set; The sample feature set is integrated and processed to obtain the full set of potential features.
6. The method according to claim 5, characterized in that, Before the encoder of the neural process network encodes the clinical sample set to obtain the sample feature set, the method further includes: The clinical sample set is divided into a support set and a target set; The initial neural process network is trained based on the support set and the target set to obtain the neural process network.
7. The method according to any one of claims 1, 5, or 6, characterized in that, Before processing the cognitive state data using a conditional variational autoencoder to obtain the clinical feature vector, the process further includes: Obtain the sample mask data corresponding to the initial sample set; The clinical sample set is obtained by augmenting the initial sample set based on the sample mask data using the conditional variational autoencoder.
8. A neurological cognitive impairment prediction system, characterized in that, include: The acquisition module is used to acquire cognitive state data; wherein, the cognitive state data includes multimodal data related to the cognitive impairment state of the patients in the neurology department; The processing module is used to encode the cognitive state data using an encoder of a variational autoencoder to obtain a clinical latent vector; wherein the clinical latent vector includes a low-dimensional embedding feature vector of the cognitive state data. The processing module is further configured to process the cognitive state data through a conditional variational autoencoder to obtain a clinical feature vector; wherein the clinical feature vector includes the pathological features of the cognitive impairment carried by the cognitive state data; The prediction module is used to predict the clinical feature vector based on the full set of latent features through the decoder of the neural process network to obtain an initial prediction result; wherein, the full set of latent features includes clinical features corresponding to a set of clinical samples with a data volume greater than a quantity threshold; The prediction module is further configured to correct the initial prediction result based on the clinical latent vector to obtain a target prediction result for the cognitive state data.