A training method and device for a cognitive impairment prediction model for ALS patients

Abstract

The present invention provides a training method and device for a cognitive impairment prediction model for ALS (amyotrophic lateral sclerosis) patients: the method comprises: obtaining cognitive assessment scale data and multimodal brain imaging data of multiple ALS patients, and marking labels to construct a training sample set; the training sample set comprises structured data after preprocessing the cognitive assessment scale data, and ALS brain imaging feature data obtained by ALS-related feature screening and feature conversion of multimodal brain imaging data; constructing a cognitive impairment prediction model, and using the data and labels in the training sample set, iteratively training the cognitive impairment prediction model to obtain a converged cognitive impairment prediction model. The present invention solves the problem that the prediction of cognitive impairment in the prior art mainly relies on cognitive scale evaluation, the examination is time-consuming and lengthy, and the influence of brain structure changes caused by ALS disease factors on cognitive impairment judgment is not considered, resulting in poor prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of clinical medical technology, and in particular relates to a training method and equipment for a cognitive impairment prediction model for ALS patients. Background Art

[0002] ALS (Amyotrophic lateral sclerosis, ALS for short), also known as motor neuron disease (commonly known as "Lou Gehrig's disease"), has a high disability and mortality rate. Recent studies have found that the proportion of ALS patients with cognitive impairment is very high, with about 10%-60% having varying degrees and forms of cognitive impairment. Cognitive dysfunction significantly affects the prognosis of ALS patients. Based on previous studies, patients with combined ALS and cognitive impairment have a faster progression of the disease and a shorter survival period than patients with simple ALS. Therefore, early detection and early treatment of ALS cognitive dysfunction are of great significance for improving the prognosis of the disease.

[0003] However, the assessment of cognitive dysfunction has always been a difficult point in clinical diagnosis and treatment. The current diagnostic standard is to conduct a comprehensive cognitive scale assessment. However, comprehensive scale examinations are time-consuming and lengthy, and are subject to strong subjective images of doctors and patients. They are difficult to carry out in clinical routine, and the sensitivity and specificity of the scales vary. The applicant's previous research has shown that due to the structural changes in the brain caused by ALS, magnetic resonance imaging indicators are correlated with clinical cognitive assessment scale scores. Radiomics is a technology based on artificial intelligence that has been developed in recent years to mine quantitative image features from standard medical images in high throughput. By combining multimodal data, it can assist in making more accurate and reliable clinical decisions.

[0004] Therefore, how to establish a cognitive impairment prediction model for ALS patients based on imaging omics and cognitive assessment scale data to improve the accuracy of predicting cognitive impairment in ALS patients has become an urgent problem to be solved. Summary of the invention

[0005] In view of the above analysis, the present invention aims to provide a training method and device for a cognitive impairment prediction model for ALS patients, which is used to solve the problem that cognitive impairment prediction in the prior art mainly relies on cognitive scale evaluation, which is time-consuming and lengthy, and does not take into account the impact of brain structure changes caused by ALS disease factors on cognitive impairment judgment, resulting in poor prediction accuracy.

[0006] The purpose of the present invention is mainly achieved through the following technical solutions:

[0007] In one aspect, the present invention provides a method for training a cognitive impairment prediction model for ALS patients, the method comprising:

[0008] Obtain cognitive assessment scale data and multimodal brain imaging data from multiple ALS patients, and annotate the corresponding patients’ cognitive impairment or non-cognitive impairment labels;

[0009] A training sample set is constructed based on the cognitive assessment scale data and the multimodal brain imaging data; the training sample set includes structured data obtained by preprocessing the cognitive assessment scale data, and ALS brain imaging feature data obtained by ALS-related feature screening and feature conversion of the multimodal brain imaging data;

[0010] Constructing a cognitive impairment prediction model, the cognitive impairment prediction model comprising an ALS brain image feature extraction module and a scale data feature extraction module that are set in parallel, and a feature fusion prediction module;

[0011] The cognitive impairment prediction model is iteratively trained using the data and labels in the training sample set to obtain a converged cognitive impairment prediction model.

[0012] Furthermore, the ALS brain image feature extraction module is obtained based on Transformer model training, and is used to extract features from the ALS brain image feature data to obtain a first feature vector;

[0013] The scale data feature extraction module is obtained based on multi-layer perceptron training and is used to extract features from the pre-processed cognitive assessment scale data to obtain a second feature vector;

[0014] The feature fusion prediction module includes a fusion layer and a fully connected layer, which are used to perform vector concatenation and classification prediction on the first feature vector and the second feature vector to obtain a cognitive impairment prediction result for ALS patients.

[0015] Furthermore, the ALS brain image feature extraction module includes 6 layers of hidden modules, the number of nodes of which are 3, 6, 12, 6, 12, and 4, respectively, and each layer of the hidden module is provided with a convolution layer, a maximum pooling layer, and a sigmoid activation layer;

[0016] The scale data feature extraction module is a two-layer multi-layer perceptron model connected in sequence, with an input channel dimension of 3, 8 neurons in the first layer, 4 neurons in the second layer, and ReLU activation functions.

[0017] Furthermore, the ALS brain image feature data is obtained by performing ALS-related feature screening and feature conversion on the multimodal brain image data by the following method, including:

[0018] The pre-processed multimodal brain imaging data are respectively annotated with ALS cognitive impairment-specific brain regions, and imaging omics features are extracted to obtain an imaging omics feature data set;

[0019] Calculate the variance between the same features in each data of the same modality in the imaging omics feature data set, reverse sort the feature data corresponding to each modality according to the variance, select the first n features with the largest variance, and divide each selected feature data by n as the sample feature data of each sample data;

[0020] Calculating the Euclidean distance between the characteristic data of each sample, using the Euclidean distance as the new characteristic data corresponding to each patient, and performing rank normalization on the new characteristic data;

[0021] The feature data after rank normalization is subjected to feature transformation by using an automatic encoder; and the feature data of different modalities after feature transformation are fused and spliced ​​to obtain ALS brain image feature data.

[0022] Furthermore, the rank normalization includes:

[0023] Each value in the new matrix of the feature data is replaced with the rank determined according to the global chain, and the rank is converted to a normal score using the inverse transformation and the score offset, expressed as:

[0024]

[0025] Among them, r (ij) is the rank of the [i,j] position in the feature data matrix, is the inverse of the matrix, and S is the number of valid samples.

[0026] Furthermore, the multimodal brain imaging data includes T1WI, T2WI, DTI and resting fMRI brain imaging data; the ALS cognitive impairment specific brain regions include:

[0027] The ALS cognitive impairment-specific brain regions corresponding to the T1WI and T2WI data include: orbitofrontal cortex, dorsolateral prefrontal lobe, entorhinal area, middle frontal gyrus, inferior frontal gyrus, precentral gyrus, superior temporal gyrus, temporal pole, insula, cingulate gyrus, striatum, thalamus, hippocampus, and amygdala;

[0028] The ALS cognitive impairment-specific brain regions corresponding to the DTI data include: corticostriatal pathway, corpus callosum corticospinal tract, cortico-brainstem tract, frontal lobe, and temporal lobe;

[0029] The ALS cognitive impairment-specific brain regions corresponding to the fMRI data include: superior frontal gyrus, middle frontal gyrus, middle temporal gyrus, precentral gyrus, postcentral gyrus, insula, lingual gyrus, and angular gyrus.

[0030] Furthermore, the cognitive assessment scale data include the Mini-Mental State Examination data MMSE and the Montreal Cognitive Assessment data MoCA, as well as the line connection test, verbal fluency test, trial recognition, prospective memory, auditory word memory test and clock drawing test scale data;

[0031] The cognitive assessment scale data were normalized and integrated to obtain structured data including Mini-Mental State Examination data (MMSE), Montreal Cognitive Assessment data (MoCA), and line connection test, verbal fluency test, trial recognition, prospective memory, auditory word memory test, and clock drawing test scale data.

[0032] Furthermore, before feature screening and feature conversion, each modal data is preprocessed by the following method:

[0033] T1WI data: Use SPM12 to convert data format, perform tissue segmentation and spatial standardization;

[0034] T2WI data: In the DICOM header window settings, convert the contrast resolution from 16 bits to 8 bits and register to the standard space;

[0035] DTI data: DICOM files were converted to different formats using dcm2nii and aligned to DTI reference data using linear registration.

[0036] Resting-state fMRI data: Data were organized temporally and spatially using dpabi.

[0037] Furthermore, the radiomics feature extraction includes:

[0038] Based on the AAL atlas, the pyradiomics library was used to extract radiomic features of specific brain regions, and 1742 features were extracted for each brain region of interest.

[0039] In another aspect, a computer device is disclosed, comprising at least one processor, and at least one memory communicatively connected to the processor;

[0040] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the aforementioned training method for the cognitive impairment prediction model for ALS patients.

[0041] Beneficial effects of this technical solution:

[0042] 1. Based on previous studies on brain region changes in patients with ALS and cognitive impairment, the present invention selects specific brain regions of interest related to ALS and cognitive impairment for feature extraction and screening, and uses sample data related to ALS and cognitive impairment to train a cognitive impairment model, thereby eliminating the influence of irrelevant feature data and improving the sensitivity, specificity and accuracy of the cognitive impairment prediction model for predicting cognitive impairment in the ALS population;

[0043] 2. The present invention combines the data of cognitive assessment scales with brain imaging data, and performs targeted screening of omics data that affect ALS cognitive impairment in the algorithm. By means of variance, Euclidean distance and rank normalization between features, features strongly correlated with ALS cognitive impairment are screened and used for training cognitive impairment models, thus obtaining a more accurate and scientific ALS cognitive impairment prediction model.

[0044] Other features and advantages of the present invention will be described in the following description, and part of them will become obvious from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0046] Figure 1 is a flow chart of a training method for a cognitive impairment prediction model for ALS patients according to an embodiment of the present invention; DETAILED DESCRIPTION

[0047] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the implementation cases of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0048] One embodiment of the present invention provides a training method for a cognitive impairment prediction model for ALS patients, such as Figure 1 As shown, the method includes:

[0049] Step S1: Acquire cognitive assessment scale data and multimodal brain imaging data of multiple ALS patients, and annotate the corresponding patients' cognitive impairment or non-cognitive impairment labels;

[0050] Specifically, multimodal brain imaging data can be acquired through a 3.0T magnetic resonance scan commonly used in medicine; the multimodal brain imaging data used in this embodiment is 4-modal brain imaging data, namely, T1 weighted phase (T1WI), T2 weighted phase (T2WI), diffusion tensor imaging (DTI) and resting-state functional magnetic resonance imaging (fMRI) in MRI (magnetic resonance imaging).

[0051] Furthermore, the cognitive assessment scale data used in this embodiment include the Mini-Mental State Examination data MMSE and the Montreal Cognitive Assessment scale data MoCA, as well as the line connection test, verbal fluency test, trial recognition, prospective memory, auditory word memory test and clock drawing test scale data;

[0052] In order to train an accurate cognitive impairment prediction model, cognitive assessment scale data and brain imaging data were collected from multiple ALS patients when constructing the sample data set, where multiple ALS patients included patients with cognitive impairment and patients without cognitive impairment. After acquiring the data, each data needs to be labeled with an accurate cognitive impairment or non-cognitive impairment label. In actual application, professionally trained specialists conduct a comprehensive cognitive scale examination on ALS patients to minimize the influence of the subjectivity of the evaluators and patients. The cognitive scale examination includes: through the sub-scale tests such as the line connection test, verbal fluency test, trial recognition, prospective memory, auditory word memory test, and clock drawing test, it is judged whether the corresponding ALS patients meet the diagnostic criteria for cognitive impairment in different aspects such as execution, language, memory, visual space, and social cognitive function; combined with the evaluation results of the mini mental status examination (MMSE) and the Montreal Cognitive Assessment (MoCA), according to the diagnostic criteria for ALS with cognitive impairment in amyotrophic lateral sclerosis-frontotemporal spectrum disease, the patients are divided into the ALS with cognitive impairment group and the ALS without cognitive impairment group; the data of patients with cognitive impairment are labeled as 1, and the data of patients without cognitive impairment are labeled as 0 for subsequent model training.

[0053] Step S2: constructing a training sample set based on the cognitive assessment scale data and the multimodal brain imaging data; the training sample set includes structured data obtained by preprocessing the cognitive assessment scale data, and ALS brain imaging feature data obtained by ALS-related feature screening and feature conversion of the multimodal brain imaging data;

[0054] Specifically, the ALS brain image feature data is obtained by performing ALS-related feature screening and feature conversion on the multimodal brain image data through the following method:

[0055] The pre-processed multimodal brain imaging data are respectively annotated with ALS cognitive impairment-specific brain regions, and imaging omics features are extracted to obtain an imaging omics feature data set;

[0056] Calculate the variance between the same features in each data of the same modality in the imaging omics feature data set, reverse sort the feature data corresponding to each modality according to the variance, select the first n features with the largest variance, and divide each selected feature data by n as the sample feature data of each sample data;

[0057] Calculating the Euclidean distance between the characteristic data of each sample, using the Euclidean distance as the new characteristic data corresponding to each patient, and performing rank normalization on the new characteristic data;

[0058] The feature data after rank normalization is subjected to feature transformation by using an automatic encoder; and the feature data of different modalities after feature transformation are fused and spliced ​​to obtain ALS brain image feature data.

[0059] Specifically, radiomics is a technology that mines quantitative image features from standard medical images at high throughput, which can be applied to clinical classification prediction systems to improve the accuracy of diagnosis, prognosis, and prediction. The implementation process includes identifying a large number of quantitative features in the image, storing the extracted data in a database (i.e., a system in which multiple independent databases operate as one entity), and then mining the data for feature extraction and application. Using radiomics data to correlate or infer outcome data can produce an accurate and reliable clinical decision-making system.

[0060] Previous studies have shown that brain structure and function tests based on imaging methods can be used to predict cognitive impairment in normal people. However, due to the changes in brain structure and function of ALS disease itself, it is necessary to establish a cognitive impairment prediction model for the ALS disease population. This embodiment is based on the study of changes in brain regions of ALS disease and cognitive impairment, and selects the features of specific regions of interest related to ALS disease and cognitive impairment to train the model, thereby improving the sensitivity and specificity of predicting cognitive impairment in the ALS population.

[0061] Preferably, before obtaining brain imaging data of each modality and performing imaging omics feature extraction, the data of each modality may be preprocessed, including:

[0062] T1WI data: Use SPM12 to convert data format, perform tissue segmentation and spatial standardization;

[0063] T2WI data: In the DICOM header window settings, convert the contrast resolution from 16 bits to 8 bits and register to the standard space;

[0064] DTI data: DICOM files were converted to DTI format using dcm2nii and aligned to DTI benchmark data (i.e., baseline images) using linear registration.

[0065] Resting-state fMRI data: Data were organized temporally and spatially using dpabi.

[0066] After preprocessing, the data of each modality are used for subsequent feature extraction and training of cognitive impairment prediction models to improve the accuracy of model prediction.

[0067] Furthermore, after preprocessing the brain imaging data of each modality, the ALS cognitive impairment-specific brain regions were first labeled; and the imaging omics features of the labeled ALS cognitive impairment-specific brain regions were extracted;

[0068] Specifically, after a large number of experimental studies, the ALS cognitive impairment-specific brain regions corresponding to the modal data used in this embodiment are as follows:

[0069] The ALS cognitive impairment-specific brain regions corresponding to the T1WI and T2WI data include: orbitofrontal cortex, dorsolateral prefrontal lobe, entorhinal area, middle frontal gyrus, inferior frontal gyrus, precentral gyrus, superior temporal gyrus, temporal pole, insula, cingulate gyrus, striatum, thalamus, hippocampus, and amygdala;

[0070] The ALS cognitive impairment-specific brain regions corresponding to the DTI data include: corticostriatal pathway, corpus callosum corticospinal tract, cortico-brainstem tract, frontal lobe, and temporal lobe;

[0071] The ALS cognitive impairment-specific brain regions corresponding to the fMRI data include: superior frontal gyrus, middle frontal gyrus, middle temporal gyrus, precentral gyrus, postcentral gyrus, insula, lingual gyrus, and angular gyrus.

[0072] After annotating the ALS-related brain regions, we can extract imaging features of specific brain regions based on the AAL atlas through the pyradiomics library, and extract 1,742 features for each brain region of interest.

[0073] Specifically, the imaging features extracted from each brain region of interest include first-order features, texture features, gray-level dependency matrix features, gray-level size region matrix features, gray-level run length matrix features, and gray-level co-occurrence matrix features, totaling 1742 features, among which:

[0074] First-order features: used to describe the distribution of voxel intensities within a specific brain region using common and basic metrics, which can be expressed as:

[0075]

[0076] Among them, Np It is expressed as the number of voxels contained in the brain region of interest, C is the optimal value used to convert the intensity to prevent negative values ​​of X(i), and X(i) represents the signal intensity of the i-th voxel in the brain region of interest;

[0077] Texture features, grayscale dependency matrix features, grayscale size region matrix features, grayscale run length matrix features and grayscale co-occurrence matrix are second-order and high-order features, among which,

[0078] Texture features: used to describe the size and shape of specific brain regions, such as voxels, surface areas, etc.

[0079] Gray-level co-occurrence matrix (GLCM): used to describe the second-order joint probability of specific brain regions;

[0080] Grayscale size zone feature matrix (GLSZM): used to quantify the grayscale area in the image. The grayscale area is defined as the number of connected voxels sharing the same grayscale intensity.

[0081] Grayscale run length matrix feature (GLRLM): defined as the number of pixels of consecutive pixels with the same gray value;

[0082] Gray-level dependency matrix feature (GLDM): represents the number of connected voxels within a specified distance that depend on the central voxel.

[0083] This embodiment extracts omics features based on the brain regions of interest related to ALS cognitive impairment corresponding to each modal data for training the cognitive impairment prediction model, thereby eliminating the influence of feature data of brain regions not related to ALS cognitive impairment, and greatly improving the accuracy of the cognitive impairment prediction model for ALS patients.

[0084] As a specific embodiment, after extracting the imaging features of multiple ALS patients, an imaging feature data set consisting of imaging feature sample data of four modalities corresponding to each patient is obtained, the variance between the same features in each data of the same modality in the imaging feature data set is calculated, the feature data corresponding to each modality are reversely sorted by variance, and the top 100 features with the largest variance in each modality data are selected; then, the 100 feature data corresponding to each modality data are divided by 100 to obtain sample feature data of each sample data.

[0085] Furthermore, the Euclidean distance between every two sample feature data is calculated, and the Euclidean distance is used as the new feature data corresponding to each sample data, and the new feature data is rank normalized by the following method:

[0086] Replace each value in the new feature data matrix with the rank determined according to the global chain, and convert the rank to a normal score using the inverse transformation and the score offset;

[0087]

[0088] Among them, r (ij) is the rank of the [i,j] position, is the inverse of the matrix, and S is the number of valid samples.

[0089] After rank normalization, each sample has The first feature has a value of 1.0 and the last feature has a value of 0.

[0090] Furthermore, for each modality of omics data, an autoencoder is used for transformation; this embodiment uses the python keras library to construct the encoder, using the Adam optimization algorithm, where the autoencoder includes 4 3x3 convolutional layers and 4 3x3 deconvolutional layers. The autoencoder undergoes nonlinear transformation to reconstruct the For the omics data of modalities, an autoencoder with a hidden layer of size h is created. The corresponding function is as follows:

[0091] f(v)=tan·h(W′·s(W·v+b)+b′);

[0092] Among them, W' and W are weight matrices, h is the size of the hidden layer, and b and b' are bias matrices.

[0093] The bottleneck layer of the autoencoder outputs the transformed features Features, you can also calculate the Pearson correlation matrix between each feature and each sample data;

[0094] Based on the correlation matrix, features with a Pearson correlation coefficient less than 0.3 were excluded. After excluding the features with a Pearson correlation coefficient less than 0.3, the feature data of the four modalities corresponding to each patient were spliced ​​and fused to obtain ALS brain image feature data for subsequent cognitive impairment prediction model training.

[0095] Furthermore, the cognitive assessment scale data used for model training include the Mini-Mental State Examination data MMSE and the Montreal Cognitive Assessment scale data MoCA, as well as the line connection test, verbal fluency test, trial recognition, prospective memory, auditory word memory test and clock drawing test scale data;

[0096] The cognitive assessment scale data are normalized and integrated to obtain structured data including the Mini-Mental State Examination data (MMSE), the Montreal Cognitive Assessment data (MoCA), and the line connection test, verbal fluency test, trial recognition, prospective memory, auditory word memory test, and clock drawing test scale data;

[0097] Based on ALS brain image feature data, structured data and corresponding labels, a training sample set is constructed.

[0098] Step S3: constructing a cognitive impairment prediction model, wherein the cognitive impairment prediction model includes an ALS brain image feature extraction module and a scale data feature extraction module that are set in parallel, and a feature fusion prediction module;

[0099] Specifically, the ALS brain image feature extraction module is a neural network model based on a Transformer model, which is used to extract features from the ALS brain image feature data to obtain a first feature vector;

[0100] The scale data feature extraction module is constructed based on a multi-layer perceptron and is used to extract features from the preprocessed cognitive assessment scale data to obtain a second feature vector;

[0101] The feature fusion prediction module includes a fusion layer and a fully connected layer, which are used to perform vector concatenation and classification prediction on the first feature vector and the second feature vector to obtain a prediction result of cognitive impairment of ALS patients;

[0102] Among them, the ALS brain image feature extraction module includes an input layer, a hidden layer and an output layer, which is used to extract features from ALS brain image feature data to obtain a 4-dimensional feature vector for subsequent feature fusion and cognitive impairment prediction.

[0103] Preferably, the hidden layer of the ALS brain image feature extraction module includes 6 hidden modules, with the number of nodes being 3, 6, 12, 6, 12, and 4 respectively, and each hidden module is set to be a convolution layer, a maximum pooling layer, and a sigmoid activation layer, and the Adam gradient descent method is used to update the model parameter weights and learning rate. Among them, the weights are initialized to random values ​​on the Gaussian distribution N(0,1), the learning rate is initially set to 0.001, and the model is iteratively trained.

[0104] The scale data feature extraction module is a two-layer MLP (multi-layer perceptron) model connected in sequence. The input channel dimension is 3, the number of neurons in the first layer is 8, the number of neurons in the second layer is 4, and the activation function is ReLU. The input structured data is sequentially extracted to obtain a 4-dimensional vector.

[0105] The four-dimensional feature vectors output by the ALS brain image feature extraction module and the scale data feature extraction module are spliced ​​together through the feature fusion evaluation module to form the input vector of the feature fusion prediction block. The feature fusion prediction module includes two fully connected layers, which use ReLU and Sigmoid activation functions respectively. After the final fully connected layer processing, the final cognitive impairment prediction result is output; the ALS brain image feature data and the cognitive assessment scale data are fused, and the prediction result representing "cognitive impairment" or "non-cognitive impairment" is output.

[0106] Step S3: using the data in the training sample set to iteratively train the cognitive impairment prediction model to obtain a converged cognitive impairment prediction model.

[0107] Specifically, the training of the cognitive impairment prediction model includes:

[0108] Setting loss functions for the ALS brain image feature extraction module, the scale data feature extraction module and the feature fusion prediction module respectively;

[0109] Assign weights to the loss function of each module to obtain the total loss function;

[0110] The data in the training sample set is used, based on the total loss function, and the gradient descent method is used to iteratively train the model parameters to obtain a converged cognitive impairment prediction model.

[0111] In this embodiment, the established cognitive impairment prediction model is trained using a loss function, and parameter training and updating are performed iteratively multiple times to reduce the error between the sample data label and the predicted value. This embodiment uses the following loss function:

[0112]

[0113] Where x represents the input sample data, y represents the actual label (cognitive impairment is 1 / non-cognitive impairment is 0), P x It represents the probability that sample x is predicted to be a positive class, and n represents the total number of samples.

[0114] After the model is trained, when the cognitive impairment prediction model is actually applied to predict cognitive impairment in ALS patients, first select 100 feature data that are the same as the sample data of the corresponding modality in the imaging genomics feature data of each modality corresponding to the patient to be evaluated, and divide each feature data by 100 to obtain new imaging genomics data; calculate the Euclidean distance between the new imaging genomics data of the patient to be evaluated and each sample data, use the Euclidean distance as the new feature data corresponding to the patient to be evaluated, and rank-normalize the new feature data and input it into the automatic encoder for feature conversion, and input the brain image feature data after feature conversion and the preprocessed cognitive assessment scale data into the trained cognitive impairment prediction model to obtain the prediction result.

[0115] In summary, the training method of the cognitive impairment prediction model for ALS patients provided by the embodiments of the present invention is based on previous studies on brain area changes in patients with ALS and cognitive impairment, selects specific brain areas of interest related to ALS and cognitive impairment for feature extraction and screening, and uses sample data related to ALS and cognitive impairment to train the cognitive impairment model, thereby eliminating the influence of irrelevant feature data and improving the sensitivity, specificity and accuracy of the cognitive impairment prediction model for predicting cognitive impairment in the ALS population; and the present invention combines data from a variety of cognitive assessment scales, and algorithmically performs targeted screening of omics data that affect ALS cognitive impairment, and screens out features that are strongly correlated with ALS cognitive impairment through methods such as variance, Euclidean distance and rank normalization between features, which are used for training the cognitive impairment model, thereby obtaining a more accurate and scientific ALS cognitive impairment prediction model.

[0116] Another embodiment of the present invention further discloses a computer device, comprising at least one processor, and at least one memory communicatively connected to the processor;

[0117] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the aforementioned training method for the cognitive impairment prediction model for ALS patients.

[0118] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0119] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A training method for a cognitive impairment prediction model for ALS patients, characterized in that: include: Obtain cognitive assessment scale data and multimodal brain imaging data from multiple ALS patients, and annotate the corresponding patients’ cognitive impairment or non-cognitive impairment labels; A training sample set is constructed based on the cognitive assessment scale data and the multimodal brain imaging data; the training sample set includes structured data obtained by preprocessing the cognitive assessment scale data, and ALS brain imaging feature data obtained by ALS-related feature screening and feature conversion of the multimodal brain imaging data by the following method: ALS cognitive impairment-specific brain area annotation is performed on the preprocessed multimodal brain imaging data, and imaging genomics feature extraction is performed to obtain an imaging genomics feature data set; the variance between the same features in each data of the same modality in the imaging genomics feature data set is calculated, the feature data corresponding to each modality are reversely sorted according to the variance, the first n features with the largest variance are selected, and each selected feature data is divided by n as the sample feature data of each sample data; the Euclidean distance between each sample feature data is calculated, the Euclidean distance is used as the new feature data corresponding to each patient, and the new feature data is rank normalized; the rank-normalized feature data is feature converted using an autoencoder; and The feature data of different modalities after feature conversion are fused and spliced ​​to obtain the ALS brain image feature data; Wherein, the rank normalization includes: Each value in the new matrix of the feature data is replaced with the rank determined according to the global chain, and the rank is converted to a normal score using the inverse transformation and the score offset, expressed as: Among them, r (ij) is the rank of the [i,j] position in the feature data matrix, is the inverse of the matrix, S is the number of valid samples; Constructing a cognitive impairment prediction model, the cognitive impairment prediction model includes an ALS brain image feature extraction module and a scale data feature extraction module, and a feature fusion prediction module, which are arranged in parallel; the ALS brain image feature extraction module is a neural network model based on a Transformer model, which is used to extract features from the ALS brain image feature data to obtain a first feature vector; the scale data feature extraction module is constructed based on a multi-layer perceptron, which is used to extract features from the preprocessed cognitive assessment scale data to obtain a second feature vector; the feature fusion prediction module includes a fusion layer and a fully connected layer, which are used to perform vector splicing and classification prediction on the first feature vector and the second feature vector to obtain a cognitive impairment prediction result for ALS patients; The cognitive impairment prediction model is iteratively trained using the data and labels in the training sample set to obtain a converged cognitive impairment prediction model.

2. The method for training a cognitive impairment prediction model for ALS patients according to claim 1, characterized in that: The ALS brain image feature extraction module includes 6 layers of hidden modules, the number of nodes are 3, 6, 12, 6, 12, 4 respectively, and each layer of hidden modules is provided with a convolution layer, a maximum pooling layer and a sigmoid activation layer; The scale data feature extraction module is a two-layer multi-layer perceptron model connected in sequence, with an input channel dimension of 3, 8 neurons in the first layer, 4 neurons in the second layer, and ReLU activation functions.

3. The method for training a cognitive impairment prediction model for ALS patients according to claim 1, characterized in that: The multimodal brain imaging data includes T1WI, T2WI, DTI and resting fMRI brain imaging data; The ALS-specific brain regions for cognitive impairment include: The ALS cognitive impairment-specific brain regions corresponding to the T1WI and T2WI data include: orbitofrontal cortex, dorsolateral prefrontal lobe, entorhinal area, middle frontal gyrus, inferior frontal gyrus, precentral gyrus, superior temporal gyrus, temporal pole, insula, cingulate gyrus, striatum, thalamus, hippocampus, and amygdala; The ALS cognitive impairment-specific brain regions corresponding to the DTI data include: corticostriatal pathway, corpus callosum corticospinal tract, cortico-brainstem tract, frontal lobe, and temporal lobe; The ALS cognitive impairment-specific brain regions corresponding to the fMRI data include: superior frontal gyrus, middle frontal gyrus, middle temporal gyrus, precentral gyrus, postcentral gyrus, insula, lingual gyrus, and angular gyrus.

4. The method for training a cognitive impairment prediction model for ALS patients according to claim 1, characterized in that: The cognitive assessment scale data include the Mini-Mental State Examination (MMSE) and the Montreal Cognitive Assessment (MoCA), as well as the line connection test, verbal fluency test, trial recognition, prospective memory, auditory word memory test and clock drawing test scale data; The cognitive assessment scale data were normalized and integrated to obtain structured data including Mini-Mental State Examination data (MMSE), Montreal Cognitive Assessment data (MoCA), and line connection test, verbal fluency test, trial recognition, prospective memory, auditory word memory test, and clock drawing test scale data.

5. The method for training a cognitive impairment prediction model for ALS patients according to claim 1, characterized in that: Before feature screening and feature conversion, each modal data is preprocessed by the following methods: T1WI data: Use SPM12 to convert data format, perform tissue segmentation and spatial standardization; T2WI data: In the DICOM header window settings, convert the contrast resolution from 16 bits to 8 bits and register to the standard space; DTI data: DICOM file format conversion is performed using dcm2nii; And linear registration is used to align to the DTI benchmark data; Resting-state fMRI data: Data were organized temporally and spatially using dpabi.

6. The method for training a cognitive impairment prediction model for ALS patients according to claim 1, characterized in that: The imaging omics feature extraction comprises: Based on the AAL atlas, the pyradiomics library was used to extract radiomic features of specific brain regions, and 1742 features were extracted for each brain region of interest.

7. A computer device, characterized in that: comprising at least one processor, and at least one memory communicatively connected to the processor; The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the training method of the cognitive impairment prediction model for ALS patients as described in any one of claims 1-6.

Images