Cognitive connection test automatic evaluation method and system and storage medium
By adopting the cognitive connection test automated evaluation method in cognitive evaluation, using electronic acquisition board and automated feature extraction module, the problem that existing evaluation methods rely on subjective judgment is solved, and the rapid and objective evaluation of cognitive impairment level is achieved, and the credibility and efficiency of the evaluation is improved.
Patent Information
- Application Number
- CN202411765247.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-04
- Publication Date
- 2025-03-11
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing cognitive assessment tests rely on subjective judgments of professional physicians, resulting in low credibility of assessment results. Manual assessment requires high labor costs and makes it difficult to record behavioral data such as reaction time.
The cognitive connection test automation evaluation method is adopted, and the subjects' trajectory data is collected through electronic acquisition boards and electromagnetic pens. The automated feature extraction module, feature fusion module and classifier module are used to achieve a rapid and objective assessment of the level of cognitive impairment.
A rapid, efficient and objective assessment of cognitive impairment levels is achieved, which reduces dependence on professional physicians, reduces assessment costs, and can record behavioral data, which improves the credibility and efficiency of assessment.
Smart Images
Figure CN119586978A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of cognitive connection testing, and in particular to an automated evaluation method, system and storage medium for cognitive connection testing. Background Art
[0002] With the development of society, the aging problem is a serious challenge facing all countries. The prevalence of age-related diseases such as Alzheimer's disease (AD) will further increase the social and economic burden. These neurodegenerative diseases can cause memory impairment, language impairment, cognitive impairment, motor imbalance and other problems in the elderly, seriously affecting their daily lives. Since degenerative diseases such as Alzheimer's disease develop gradually, if mild cognitive impairment (MCI) can be detected and intervened in time, timely medical care can be provided, thereby delaying the formation of Alzheimer's disease.
[0003] Cognitive assessment tests such as the Mini-Mental State Examination (MMSE) and the Montreal Cognitive Assessment (MoCA) are relatively simple and easy to obtain ways to quickly assess cognitive impairment, but their assessment methods mainly rely on the subjective judgment of professional physicians, which introduces the subjective bias of the raters. Therefore, the credibility of the assessment results is not high and the assessment effect is not good. On the other hand, manual assessment not only requires high labor costs, but also makes it difficult to record behavioral data such as reaction time. Summary of the invention
[0004] 1. Technical issues to be resolved In view of the deficiencies in the prior art, the present invention provides an automated evaluation method, system and storage medium for a cognitive connection test, which have the advantages of being able to quickly, efficiently and objectively evaluate the cognitive impairment level of a subject.
[0005] (II) Technical solution In order to achieve the above-mentioned purpose of being able to quickly, efficiently and objectively evaluate the cognitive impairment level of the subject, the present invention provides the following technical solution: a method for automated evaluation of cognitive connection test, comprising the following steps: S1. Lay the paper TMT scale flat on the electronic collection board and align it with the anchor points marked on the electronic collection board; S2, the subjects followed the written instructions and used the electromagnetic pen to complete the TMT connection task; S3, obtaining the trajectory data of the subject; S4, extracting features from the trajectory data through an automated feature extraction module to obtain a low-dimensional feature vector, and calculating the similarity between the feature vector and the clustering template and the template encoding ID; S5, obtaining relevant features of the TMT task through a relevant feature extraction module; S6, concatenating and fusing the above extracted features through a feature fusion module; S7. Input the concatenated and fused features into the classifier module to obtain the assessment results of the cognitive impairment level of the subjects.
[0006] As a preferred technical solution of the present invention, the trajectory data includes pen touch point and pressing pressure data, which are collected at a fixed frequency through an electronic collection board.
[0007] As a preferred technical solution of the present invention, it also includes a data preprocessing module for preprocessing the trajectory data of the subject; Preprocessing includes: data smoothing, length padding, and removing redundant data.
[0008] As a preferred technical solution of the present invention, the automatic feature extraction module extracts features from trajectory data based on a semi-supervised vector quantization variational autoencoder.
[0009] As a preferred technical solution of the present invention, the relevant feature extraction module includes a time feature extraction module, a pressure feature extraction module, a smoothness feature extraction module and a statistical feature query module.
[0010] As a preferred technical solution of the present invention, the time feature extraction module is used to calculate the completion time and preparation time of the TMT task; The pressure feature extraction module is used to calculate the pressure features during the execution of the task; The smoothness feature extraction module is used to calculate the smoothness metric of the trajectory; The statistical feature query module is used to query statistical features in the medical record system according to the subject's ID number, and the statistical features include gender, age and education level.
[0011] As a preferred technical solution of the present invention, the feature fusion module uses principal component analysis or independent component analysis method to reduce the dimension of the feature vector.
[0012] As a preferred technical solution of the present invention, the classifier module includes a random forest classifier, a bagging classifier, an AdaBoost classifier and an XGBoost classifier.
[0013] (III) Beneficial effects Compared with the prior art, the present invention provides an automated evaluation method for cognitive connection testing, which has the following beneficial effects: 1. The automated assessment method for cognitive connection test, through automated feature extraction and analysis of the collected cognitive connection test data, does not need to rely on the assessment of professionally trained psychiatrists, and can accurately and quickly output the assessment results of the level of cognitive impairment, thus realizing objective and reproducible automated auxiliary diagnosis of the level of cognitive impairment.
[0014] 2. The automated evaluation method of the cognitive connection test combines traditional statistical features (such as gender, age, and education level) and highly interpretable computer-extracted features (such as time features, pressure features, smoothness features, and low-dimensional latent expression features based on semi-supervised vector quantization variational autoencoders) in the feature extraction stage. It is more understandable and acceptable to clinicians and is of great help in auxiliary diagnosis.
[0015] 3. This automated assessment method for cognitive connection testing does not require professionally trained doctors to interpret and judge the results of TMT tasks. It only relies on computer computing power to achieve efficient and simple assessment, reducing the waiting time of the subjects. It provides a promising solution for large-scale community screening and home telemedicine, which is of great significance for the early screening and treatment of Alzheimer's disease. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 is an overall flow chart of the evaluation method of the present invention; Figure 2 This is a neural network architecture diagram of the semi-supervised vector quantization variational autoencoder of the present invention. DETAILED DESCRIPTION
[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0018] Embodiment 1: See also Figure 1 and Figure 2 , a cognitive connection test automated evaluation method, comprising the following steps: S1. Lay the paper TMT scale flat on the electronic collection board and align it with the anchor points marked on the electronic collection board; S2. Subjects follow written instructions and use electromagnetic pens to complete the TMT connection task, which includes: Task type 1: Alternately connect the numbers surrounded by circles and squares in order from small to large; Task type 2: Add additional squares or circles of the same number around the circle or square on the correct connection track, except for the first number.
[0019] S3. Acquire the subject's trajectory data, which includes pen touch point and pressing pressure data. The data is collected at a fixed frequency by an electronic collection board, and the collected data is stored in a data storage module.
[0020] S4. Preprocess the trajectory data of the subjects through the data preprocessing module. The preprocessing includes: data smoothing, length padding, and removal of redundant data. The smoothing methods of the reference trajectory include convolution smoothing, sliding average smoothing, and B-spline smoothing; the length padding methods include filling the end with 0, filling the end with the value of the last sampling point, and filling the end with the trajectory average. When padding, the pressure value is filled with 0; at the same time, since some subjects cannot complete the test, the data is empty, and the padding methods include filling the end with 0, -1, etc.; removing redundant data is achieved by equidistant downsampling the TMT collected data to remove redundant data, reduce the computing resource loss of model training, and reduce the time of model inference.
[0021] Then, the trajectory data is extracted through the automatic feature extraction module to obtain a low-dimensional feature vector, and the similarity between the feature vector and the clustering template and the template encoding ID are calculated. Specifically: The automatic feature extraction module based on the semi-supervised vector quantization variational autoencoder includes a pair of encoding networks and decoding networks and a set of trainable one-dimensional tensors (or template sets), wherein the encoding network is used to extract the low-dimensional expression features of the collected TMT trajectory data; then the distance between the low-dimensional expression features and each one-dimensional tensor in the template set is calculated respectively, and the tensor distances and tensor numbers of the template set are arranged in order from small to large or from large to small, which together constitute the template features; then the low-dimensional expression features output by the encoding network are concatenated with the template features obtained by the above process as the final encoding features, and the decoding network is only used in the model training stage to supervise whether the low-dimensional expression features extracted by the encoding network are representative. Optionally, the above-mentioned distance calculation methods include Euclidean distance, Manhattan distance, cosine distance, etc.
[0022] S5, obtaining relevant features of the TMT task through a relevant feature extraction module; The relevant feature extraction modules include time feature extraction module, pressure feature extraction module, smoothness feature extraction module and statistical feature query module. Specifically: The time feature extraction module is used to calculate the completion time and preparation time of the TMT task (or thinking time, which can be counted based on the number of points where the recorded stress is less than or equal to 0); The pressure feature extraction module is used to calculate the pressure features during the task execution (including minimum pressure, maximum pressure, and average pressure); The smoothness feature extraction module is used to calculate the smoothness metric of the trajectory (the sum of the squares of the third-order acceleration); The statistical feature query module is used to query statistical features in the medical record system according to the subject's ID number. The statistical features include gender, age and education level; Of the above related features, you can choose to use only some of them as needed.
[0023] S6. The above extracted features are concatenated and fused through the feature fusion module. At the same time, the principal component analysis (PCA) or independent component analysis (ICA) method can be used to reduce the dimension of the feature vector.
[0024] S7. Input the concatenated and fused features into the classifier module to obtain the cognitive impairment level assessment results of the subjects. The classifier module specifically includes random forest classifier, bagging classifier, AdaBoost and XGBoost classifier.
[0025] In this embodiment, the neural network architecture diagram of the semi-supervised vector quantization variational autoencoder can be referred to Figure 2 The semi-supervised vector quantization variational autoencoder architecture mainly includes a pair of encoding networks and decoding networks and a set of trainable one-dimensional tensors (or template sets). The encoding network is used to extract the low-dimensional expression features of the collected TMT trajectory data. The middle layer of the encoding network consists of multiple one-dimensional convolution and pooling layers. Optional, the pooling layer types include average pooling and maximum pooling; The last layer of the encoding network is a fully connected layer; then the distance between the low-dimensional expression feature output by the encoding network and each one-dimensional tensor in the template set is calculated respectively, and the tensor distances and tensor numbers of the template set are arranged in order from small to large or from large to small, which together constitute the template feature. Optionally, the distance calculation method includes Euclidean distance, Manhattan distance, cosine distance, etc.
[0026] The role of the template set here is to form multiple representative templates (or reference templates) that have significant correlations with the normal control group, MCI or AD. During the training process, the distance loss function between the defined features and the template will make the representative template gradually have similar representations to multiple low-dimensional expression features in the group when the back propagation is performed; the sample proportion loss function of the defined template representing the same cognitive level will make the representative template more representative and far away from the low-dimensional expression features of other cognitive categories when the back propagation is performed. Then, the low-dimensional expression features output by the encoding network are spliced with the calculated template features as the final encoding features. Finally, the low-dimensional expression features output by the encoding network are spliced with the template closest to it in the template set, and input into the decoding network to obtain the decoded TMT trajectory data. The decoding network is composed of fully connected layers. In fact, the decoding network does not participate in the calculation in the inference stage. Its function is to supervise whether the low-dimensional expression features extracted by the encoding network are representative during the model training stage.
[0027] The loss function definition of the neural network of the semi-supervised vector quantization variational autoencoder in the present invention consists of three parts: (1) First, the reconstructed distance error between the collected real cognitive connection test TMT trajectory data and the decoded TMT trajectory data. The distance function here can be Euclidean distance or Manhattan distance.
[0028] (2) The second is the distance error between the low-dimensional expression features output by the encoding network and the template closest to it in the template set. The distance function here can be Euclidean distance, Manhattan distance, cosine distance, etc.
[0029] (3) Finally, the proportion of samples with the same cognitive level represented by the template is the largest. First, each template in the template set is classified, and the classification rule is: 1. Find the nearest template for each sample in the input set and match them one by one. Then count the categories of all samples corresponding to each template, and take the category with the most samples belonging to the normal control group, MCI or AD as the category of the template. In this way, each template is labeled with the label of the normal control group, MCI or AD.
[0030] 2. Then calculate the ratio of the number of samples with the same template category corresponding to each template to the samples with the same template category in the input set, and calculate the ratio of the number of samples of the corresponding template category to all samples corresponding to the template, and multiply and invert the two ratios.
[0031] The functions of the three parts defined by the loss function are: (1) The reconstruction distance loss supervises whether the low-dimensional expression features extracted by the encoding network are representative during the model training stage. If the decoding output of the decoding network has the same distribution as the true input sample, it means that the low-dimensional expression features are representative.
[0032] (2) The distance loss between the expression feature and the template will cause the representative template to gradually have similar representations to multiple low-dimensional expression features in the group, so the representative template has stronger group representativeness.
[0033] (3) The loss of the proportion of samples with the same cognitive level represented by the template will make the representative template more representative and away from the low-dimensional expression characteristics of other cognitive categories.
[0034] Embodiment 2: This embodiment provides a TMT-based cognitive impairment level assessment system, including: Electronic collection equipment, including an electronic collection board and a matching electromagnetic pen, used to collect the pen touch point and pressing pressure data of the subjects in the TMT task; A data storage module, used to receive and store trajectory data collected by the electronic collection device; The automatic feature extraction module is used to extract features from trajectory data, obtain low-dimensional feature vectors, and calculate the similarity between the feature vectors and the clustering template and the template encoding ID; Relevant feature extraction module, used to obtain relevant features of TMT tasks; Feature fusion module, used to splice and fuse the extracted features; The classifier module is used to output the cognitive impairment level assessment result through the classifier based on the spliced and fused features.
[0035] Embodiment three: This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, an automated evaluation method for cognitive connection testing as described in the first embodiment is implemented.
[0036] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A cognitive connection test automated evaluation method, characterized in that: The following steps are involved: S1. Lay the paper TMT scale flat on the electronic collection board and align it with the anchor points marked on the electronic collection board; S2, the subjects followed the written instructions and used the electromagnetic pen to complete the TMT connection task; S3, obtaining the trajectory data of the subject; S4, extracting features from the trajectory data through an automated feature extraction module to obtain a low-dimensional feature vector, and calculating the similarity between the feature vector and the clustering template and the template encoding ID; S5, obtaining relevant features of the TMT task through a relevant feature extraction module; S6, concatenating and fusing the above extracted features through a feature fusion module; S7. Input the concatenated and fused features into the classifier module to obtain the assessment results of the cognitive impairment level of the subjects.
2. The method for automated evaluation of cognitive connection test according to claim 1, characterized in that: The trajectory data includes pen touch point and pressing pressure data, which are collected at a fixed frequency through an electronic collection board.
3. A cognitive connection test automated evaluation method according to claim 1 or 2, characterized in that: It also includes a data preprocessing module for preprocessing the trajectory data of the subject; Preprocessing includes: data smoothing, length padding, and removing redundant data.
4. The method for automated evaluation of cognitive connection test according to claim 1, characterized in that: The automatic feature extraction module extracts features from trajectory data based on a semi-supervised vector quantization variational autoencoder.
5. The method for automated evaluation of cognitive connection test according to claim 1, characterized in that: The related feature extraction module includes a time feature extraction module, a pressure feature extraction module, a smoothness feature extraction module and a statistical feature query module.
6. The method for automated evaluation of cognitive connection test according to claim 5, characterized in that: The time feature extraction module is used to calculate the completion time and preparation time of the TMT task; The pressure feature extraction module is used to calculate the pressure features during the execution of the task; The smoothness feature extraction module is used to calculate the smoothness metric of the trajectory; The statistical feature query module is used to query statistical features in the medical record system according to the subject's ID number, and the statistical features include gender, age and education level.
7. The method for automated evaluation of cognitive connection test according to claim 1, characterized in that: The feature fusion module uses principal component analysis or independent component analysis method to reduce the dimension of the feature vector.
8. The method for automated evaluation of cognitive connection test according to claim 1, characterized in that: The classifier modules include random forest classifier, bagging classifier, AdaBoost and XGBoost classifier.
9. A TMT-based cognitive impairment level assessment system, characterized in that: include: Electronic collection equipment, including an electronic collection board and a matching electromagnetic pen, used to collect the pen touch point and pressing pressure data of the subjects in the TMT task; A data storage module, used to receive and store trajectory data collected by the electronic collection device; The automatic feature extraction module is used to extract features from trajectory data, obtain low-dimensional feature vectors, and calculate the similarity between the feature vectors and the clustering template and the template encoding ID; Relevant feature extraction module, used to obtain relevant features of TMT tasks; Feature fusion module, used to splice and fuse the extracted features; The classifier module is used to output the cognitive impairment level assessment result through the classifier based on the spliced and fused features.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the automated evaluation method for cognitive connection testing as described in any one of claims 1 to 8 is implemented.