Screening and evaluating system and method for cognitive impairment of old people

Abstract

The invention discloses a screening and evaluating system and method for cognitive impairment of old people, and belongs to the technical field of intelligent medical treatment. According to the system, electroencephalogram, VR / AR interaction, physiology and gait data are collected through the multi-modal acquisition module, the data processing module carries out data processing, dynamically selects a scale and carries out fusion analysis, the man-machine interaction module realizes natural interaction, and the output module generates a visual report and carries out early warning. The core algorithm comprises dynamic scale selection, multi-modal fusion evaluation and the like. The evaluation method comprises the steps of data acquisition, scale selection, data fusion, confidence verification, output and the like. Compared with a traditional method, the method overcomes the defects that subjectivity is high, and static evaluation, data splitting and base layer application are limited. The cognitive impairment assessment method can accurately and dynamically assess cognitive impairment, improves assessment accuracy, realizes full utilization of data, is simple and convenient to operate, is suitable for various scenes, provides efficient and reliable support for diagnosis, monitoring and intervention of cognitive impairment, and promotes development of intelligent medical treatment.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent healthcare, and particularly relates to a screening and evaluation system and method for elderly cognitive impairment. Background Art

[0002] With the intensification of the global population aging, cognitive impairment has become a major public health challenge. According to the data of the National Bureau of Statistics in 2023, the population aged 60 and above in China accounts for 19.84%. The research in The Lancet Public Health shows that in 2020, the number of patients with mild cognitive impairment (MCI) aged 60 and above in China reached 38.77 million, with a prevalence rate of 15.5%. Cognitive impairment not only seriously affects the quality of life of patients, but also brings a heavy burden to families and society.

[0003] There are many defects in the existing technology for cognitive impairment evaluation:

[0004] Strong subjectivity: Traditional scales (such as MMSE, MoCA) rely on the subjective answers of patients and are significantly affected by educational level and emotional state, with an error rate as high as 15%-20%. This makes the evaluation results may not accurately reflect the true cognitive state of patients and affects the accuracy of diagnosis.

[0005] Static evaluation: Single detection cannot capture the dynamic decline trajectory of cognitive function (such as the MMSE score drops by 0.5-1 point per month in the early stage of AD). It is difficult for doctors to comprehensively understand the disease development of patients, which is not conducive to formulating personalized treatment plans and timely adjusting intervention measures.

[0006] Data fragmentation: Scale data, wearable device data, and hospital imaging / gene data lack a unified analysis platform, and the utilization rate is less than 30%. Data from different sources cannot be mutually verified and integrated, resulting in a waste of information resources and reducing the accuracy and efficiency of evaluation.

[0007] Limited application at the grass-roots level: Traditional evaluation requires more than 30 minutes by professional personnel and is difficult to popularize in grass-roots hospitals. This makes a large number of potential patients unable to be screened and diagnosed in time, delaying the treatment opportunity and also restricting the wide development of cognitive impairment prevention and control work. Summary of the Invention

[0008] In view of the above defects of the existing technology, the technical problem to be solved by the present invention is to provide a screening and evaluation system and method for elderly cognitive impairment, which can relatively accurately conduct dynamic evaluation and precise evaluation of patients' cognitive impairment.

[0009] To achieve the above object, the present invention provides a screening and evaluation system for elderly cognitive impairment, including:

[0010] A multimodal acquisition module for acquiring electroencephalogram signals, virtual reality (VR) / augmented reality (AR) interaction behavior data, gait parameters, and neurotransmitter level data;

[0011] A data processing module, including:

[0012] A dynamic scale selection engine, based on rule engine preliminary screening and random forest model optimization, dynamically matches the optimal scale combination according to patient characteristics;

[0013] A multimodal fusion model, which weights and fuses the standardized scores of the scale (Z scale ), electroencephalogram theta / beta power ratio (R θ / β), gait coefficient of variation (CV gait ), standardized hippocampal volume (V hippo ), and comprehensive neurotransmitter index (N) through the entropy weight method, outputs the cognitive impairment risk value, and dynamically adjusts the weights through the entropy weight method;

[0014] A confidence verification module, which calculates the internal consistency, multimodal correlation, and hospital data support degree of the evaluation result. The formula is:

[0015] C = 0.35·α + 0.25(|r EEG | + |r gait | + |r N |) + 0.4·I hospital

[0016] α is the internal consistency of the scale, r EEG is the correlation coefficient between the total score of the scale and the theta / beta power ratio, r gait is the correlation coefficient between the total score of the scale and gait parameters, r N is the correlation coefficient between the total score of the scale and neurotransmitter levels, I hospital is the hospital data support degree;

[0017] A human-computer interaction module, including a voice assistant and VR / AR devices, dynamically adjusts the task difficulty and identifies the user's emotions.

[0018] The present invention also discloses a screening and evaluation method for elderly cognitive impairment, including the following steps:

[0019] S100. Data acquisition step:

[0020] S110. Acquire resting-state and task-state electroencephalogram signals through an electroencephalogram cap;

[0021] S120. Execute a virtual supermarket task through VR / AR devices, and record the task completion time, number of wrong selections, and path efficiency;

[0022] S130. Acquire the gait coefficient of variation and plantar pressure entropy through a smart bracelet and piezoelectric insoles;

[0023] S140. Collect neurotransmitter level data and hospital imaging / gene data;

[0024] S200. Dynamic scale selection steps:

[0025] S210. Based on the patient's age, years of education, APOE genotype, gait parameters, and neurotransmitter levels, initially screen the scale range through a rule engine;

[0026] S220. Use a random forest model to calculate the scale applicability score and optimize the selection of the optimal scale combination;

[0027] S300. Data processing and fusion steps:

[0028] S310. Clean, normalize, and standardize the multimodal data;

[0029] S320. Dynamically weight and fuse the scale scores, EEG features, gait parameters, hippocampal volume, and neurotransmitter indicators through the entropy weight method to calculate the cognitive impairment risk value;

[0030] S400. Confidence verification steps:

[0031] S410. Calculate the internal consistency (Cronbach's α), multimodal data correlation (r EEG , r gait , r N ) and hospital data support (I hospital ) of the evaluation results to generate a comprehensive confidence value;

[0032] S420. Adjust the evaluation strategy according to the confidence value: if C < 0.6, supplement data and replace the scale; if 0.6 ≤ C < 0.8, increase the task difficulty; if C ≥ 0.8, generate the final report;

[0033] S500. Output and dynamic monitoring steps:

[0034] S510. Generate a multimodal report including a cognitive trend line chart and voice broadcast;

[0035] S520. Real-time monitor data changes through the cloud platform and trigger warning signals.

[0036] The beneficial effects of the present invention are:

[0037] The screening and evaluation system and method for elderly cognitive impairment of the present invention effectively overcome the defects of the prior art, have significant advantages in the field of cognitive impairment evaluation, and are of great significance for promoting the development of intelligent medicine. Specifically, it has the following advantages:

[0038] 1. High assessment accuracy: The error rate of traditional scale assessment is as high as 15%-20%, while the present invention collects multi-source data such as EEG, VR / AR interaction, physiology and gait through a multimodal acquisition module to fully reflect the patient's cognitive state. The multimodal fusion model uses the entropy weight method to dynamically weight and fuse various types of data to accurately calculate the risk value of cognitive impairment. For example, in the assessment of patients in the geriatric department of a general hospital, the assessment accuracy after multimodal fusion is over 80%, far exceeding the assessment accuracy of single-modal data. It can accurately judge the type and degree of cognitive impairment, and provide a key basis for doctors to formulate precise treatment plans.

[0039] 2. Realize dynamic monitoring: Most existing technologies are static assessments that cannot capture dynamic changes in cognitive function. The present invention uses a cloud platform and a time series analysis model to upload and analyze multimodal data of patients in real time, and promptly detect changes in cognitive status and issue warnings. Taking the health management of the elderly in nursing homes as an example, through long-term monitoring of VR / AR cognitive tests, EEG and gait data, it is possible to keenly detect subtle changes in the cognitive function of the elderly, such as finding that the reaction time of a certain elderly person's cognitive test is prolonged, the coefficient of variation of the step length is increased, etc., and timely arrange personalized training and care to effectively delay the progression of the disease.

[0040] 3. Full use of data: In the past, scales, wearable devices and hospital data were separated from each other, and the utilization rate was less than 30%. The present invention builds a unified analysis platform, integrates multi-source data, and mines potential correlations between data. In the follow-up evaluation of patients in rehabilitation institutions, various types of data in rehabilitation training are integrated to comprehensively and accurately evaluate the cognitive improvement of patients, provide a scientific basis for adjusting the rehabilitation plan, and improve the rehabilitation effect.

[0041] 4. Convenient application at the grassroots level: Traditional assessments require professionals to spend more than 30 minutes, which is difficult to carry out in grassroots hospitals. The present invention adopts lightweight interactive design, such as mobile phone AR scenes and voice assistants, which are easy to operate. Combined with the edge computing module, data can be processed quickly on grassroots devices, reducing dependence on professionals and high-end equipment. In community cognitive impairment screening, with the help of simple equipment and optimized processes, large-scale screening can be completed efficiently, and potential patients can be identified and referred in a timely manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the system architecture of the present invention;

[0043] Figure 2 It is a schematic diagram of the operation flow of the present invention. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0045] The implementation process of this embodiment is as follows:

[0046] (1) System setup

[0047] Hardware device deployment: Deploy corresponding hardware devices in application sites such as general hospitals, community service centers, rehabilitation institutions, nursing homes, and research institutions. Install the Emotiv EPOC+ EEG cap to ensure good contact between the electrodes and the scalp for accurate acquisition of EEG signals. Equip VR devices such as Oculus Quest 3 and mobile devices with AR functions (such as mobile phones supporting ARCore or ARKit), and ensure their stable performance and smooth operation of corresponding VR / AR cognitive test programs. Wear smart bracelets and piezoelectric insoles for the subjects to ensure their normal operation and accurate acquisition of gait data. At the same time, equip professional neurotransmitter detection equipment such as high-performance liquid chromatography and enzyme-linked immunosorbent assay in the laboratory to detect neurotransmitter levels.

[0048] Software system installation and configuration: Install and configure data processing software on data processing terminals (such as servers or high-performance computers). Deploy relevant algorithms and programs such as a dynamic scale selection engine, a multi-modal fusion model, a confidence verification module, and an edge computing module. Ensure smooth data transmission and interaction between software modules, and at the same time establish effective connections with hardware devices to achieve real-time acquisition, processing, and analysis of data. Build and configure the cloud platform to ensure its powerful data storage, computing, and analysis capabilities to support real-time upload, processing, and sharing of multi-modal data, and achieve the dynamic monitoring and warning functions of the cloud platform.

[0049] Network environment setup: Build a stable network environment to ensure the timeliness and security of data transmission. Inside the application site, establish a high-speed local area network to ensure stable data transmission between each hardware device and the data processing terminal. Connect the data processing terminal to the cloud platform through the Internet, and adopt the dual authentication mechanism of HTTPS+OAuth2.0 to ensure the security of data during network transmission and prevent data leakage and tampering.

[0050] (2) System usage process

[0051] Subject preparation: Before conducting a cognitive impairment assessment, introduce the assessment process and precautions to the subjects in detail to ensure that the subjects understand and cooperate with the assessment process. Wear the EEG cap, smart bracelet, and piezoelectric insoles for the subjects, and guide the subjects to correctly use VR / AR devices. In cases where blood or cerebrospinal fluid samples need to be collected to detect neurotransmitter levels, collect samples according to the standardized medical operation process.

[0052] Data collection start: The subject starts the assessment program. First, the system interacts with the subject through a voice assistant (such as "Doubao, start the cognitive test") to guide the subject to complete various cognitive function tests. During the cognitive function data collection stage, the Auditory Verbal Learning Test, the Digit Cancellation Test, part B of the Trail Making Test, the Boston Naming Test, the Clock Drawing Test, etc. are carried out in sequence, and the system automatically records the relevant data. At the same time, electroencephalogram (EEG) activity data, including resting-state and task-state EEG data, are collected; brain imaging data are obtained through devices such as MRI and PET; gait data are collected using smart bracelets and piezoelectric insoles; neurotransmitter level data are obtained through laboratory tests.

[0053] Scale selection and data processing: The dynamic scale selection engine determines the optimal scale combination according to the subject's feature vectors (such as age, years of education, APOE genotype, gait CV Step , cognitive function parameters, neurotransmitter levels, etc.) through the initial screening of the rule engine and the optimization of the random forest model. The collected data is transmitted to the data processing module for preprocessing operations such as data cleaning and normalization to remove noise and outliers and unify the data format and scale. Then, a multi-modal fusion model is used for weighted fusion to calculate the cognitive impairment risk value. During the fusion process, according to the characteristics and importance of different data, the weights are dynamically adjusted through the entropy weight method to give full play to the advantages of each data and achieve accurate assessment of cognitive impairment. At the same time, the edge computing module processes the EEG signals in real time and only uploads the key features to the cloud, reducing the data transmission volume and improving the processing efficiency.

[0054] Human-computer interaction and real-time feedback: During the assessment process, the subject participates in cognitive tasks through VR / AR devices, and the system interacts with the subject in real time through a voice assistant. The voice assistant dynamically adjusts the question difficulty according to the subject's performance in the task. For example, when the error rate > 20% is detected, the task complexity is reduced. At the same time, the system collects the subject's behavioral data and voice data in real time, and uses the BERT model to identify the user's anxiety emotion. If anxiety emotion is detected, a comforting voice ("Don't worry, take your time") is triggered to improve the user experience. In the VR / AR task, the system accurately captures the user's gestures and product selection behaviors. For example, in the "AR virtual supermarket", the executive function is evaluated according to data such as the speed at which the user searches for products and the number of incorrect selections.

[0055] Confidence Verification and Report Generation: The confidence verification module calculates the confidence of the evaluation results and determines subsequent operations based on the confidence value. If the confidence meets the requirements (e.g., C≥0.8), the system calls the Doubao template engine to automatically generate a visual report. The report content includes the cognitive state evaluation results, the detailed performance of each cognitive domain, the comparative analysis with the same-age population, and the recommended intervention measures, etc. The report supports voice broadcast ("Your memory score has increased by 10% compared to last month"), which is convenient for users to understand the evaluation results. If the confidence does not meet the requirements (e.g., C<0.6), the system replaces the scale combination and supplements hospital examinations to obtain more accurate data; when 0.6≤C<0.8, the VR / AR task difficulty is increased to further examine the patient's cognitive ability, and then the evaluation is carried out again until a reliable evaluation result is achieved.

[0056] Cloud Platform Monitoring and Early Warning: The multi-modal data generated during the evaluation process is uploaded to the cloud platform in real time. The cloud platform uses time series analysis models and machine learning algorithms to deeply analyze the data and timely detect changes in the patient's cognitive state. When it is monitored that the patient's cognitive function indicators show an obvious downward trend and exceed the preset early warning value, the system automatically sends early warning information to the doctor and the patient's family members to remind them to take intervention measures in a timely manner. Doctors can view the patient's historical data and evaluation results through the cloud platform at any time for remote diagnosis and treatment plan adjustment.

[0057] (III) System Maintenance and Optimization

[0058] Equipment Maintenance: Regularly calibrate and maintain the data collection equipment (such as EEG caps, VR devices, smart bracelets, piezoelectric insoles, etc.) to ensure the accuracy and stability of the equipment. Check the conductivity of the EEG cap electrodes and replace the damaged electrodes in a timely manner; calibrate the sensors of the VR device to ensure its tracking accuracy; regularly check the battery power and sensor performance of the smart bracelet and piezoelectric insole to ensure the reliability of data collection. Regularly maintain and calibrate the laboratory testing equipment to ensure the accuracy of neurotransmitter detection results.

[0059] Data Management: Establish a perfect data management mechanism to classify, store, and back up the collected data to ensure the security and integrity of the data. Regularly clean and update the data, delete expired or useless data, and improve the data storage and processing efficiency. At the same time, strictly abide by data privacy protection regulations, adopt differential privacy mechanisms for sensitive data such as gait and EEG, and add Gaussian noise (ε = 0.1) for privacy protection to ensure the security of the data during analysis and sharing.

[0060] Algorithm Optimization: With the development of technology and the accumulation of data, regularly optimize and update the algorithms in the system (such as dynamic scale selection algorithm, multi-modal fusion evaluation model, confidence verification mechanism, lightweight EEGNet model, etc.). By collecting more clinical data and feedback information, improve the performance and accuracy of the algorithms. Adopt new machine learning technologies and methods to enhance the generalization ability and adaptability of the models, so as to better meet the cognitive impairment assessment needs in different scenarios.

[0061] User Feedback Collection and Improvement: Actively collect feedback from users such as medical institutions, community service centers, rehabilitation institutions, nursing homes, and research institutions, and understand the problems and deficiencies existing in the actual application of the system. According to user feedback, optimize and improve the functions, interface design, operation process, etc. of the system, improve the usability and practicality of the system, and further enhance the user experience.

[0062] Through the above systematic and comprehensive implementation methods, ensure that the screening and evaluation system and method for elderly cognitive impairment provided by the present invention can operate effectively in different scenarios, provide reliable support for the diagnosis, monitoring, and intervention of cognitive impairment, and promote the development and progress of the intelligent medical field in cognitive impairment assessment.

[0063] (IV) Data Synchronization and Privacy Protection

[0064] Differential Privacy Mechanism: Design a differential privacy mechanism to add Gaussian noise to sensitive data such as gait and EEG:

[0065] D noisy = D rure + N(0, σ 2 )

[0066] D noisy represents the data after adding noise. In the scenario of data synchronization and privacy protection, this is the processed data version used for analysis and sharing, aiming to protect user privacy.

[0067] D rure represents the original data. This is the directly collected data without processing, which contains users' sensitive information.

[0068] N(0, σ 2 ) represents Gaussian noise with a mean of 0 and a variance of σ 2 . This is the noise added to the original data to obfuscate the data and thus protect user privacy. Among them, σ is the standard deviation of the noise, which determines the size and intensity of the noise.

[0069] In practical applications, it is necessary to weigh and select an appropriate σ value (or ∈ value) according to privacy protection requirements and data sensitivity. In the present invention, ε = 0.1.

[0070] In the differential privacy mechanism, the selection of values is a process of balancing the privacy protection strength and data availability. Generally, smaller values provide stronger privacy protection but have a greater impact on data availability; larger values reduce the privacy protection strength to a certain extent but can retain more data authenticity.

[0071] The selection of ε = 0.1 is based on the following considerations: On the one hand, under the premise of meeting certain privacy protection requirements, this value has relatively little perturbation to the original data, ensuring that the fused data still has high availability in the cognitive impairment assessment model and preventing the data from being unable to effectively reflect the patient's true cognitive state due to excessive perturbation. On the other hand, in the research and practice of related fields, in similar medical data privacy protection scenarios, values in the range of 0.05 - 0.2 can achieve a better balance between privacy protection and data availability. This value has been proven to effectively protect the privacy of users' sensitive data while meeting the requirements of relevant regulations such as GDPR / the Personal Information Protection Law.

[0072] Secure transmission and authentication: The data is encrypted and transmitted through the built-in security chip of Doubao, and HTTPS + OAuth2.0 dual authentication is adopted to ensure the security of data transmission and the legitimacy of user identities.

[0073] (V) System accuracy verification method

[0074] Data sources: A wide range of multi-source data is collected. In terms of clinical data, the cognitive assessment results of patients are obtained, covering various scale scores such as MoCA and MMSE; multi-modal data includes electroencephalogram data, imaging data (such as brain CT and MRI images), physiological parameters (such as heart rate, gait-related parameters collected by smart bracelets and piezoelectric insoles, such as coefficient of variation of step length, plantar pressure entropy, etc.), and neurotransmitter level data. At the same time, based on comprehensive examination results (including MRI, genetic testing, etc.), clinicians give accurate cognitive impairment diagnosis results as the true labels.

[0075] Data division: The dataset is scientifically and reasonably divided into a training set, a validation set, and a test set. Usually, a division method with 70% for training, 15% for validation, and 15% for testing can be adopted. Such a division helps to fully test the model performance at different stages and ensure the generalization ability and accuracy of the model.

[0076] The specific implementation manner of this embodiment:

[0077] Combined with Figure 1 , the screening and assessment system for elderly cognitive impairment in this embodiment includes:

[0078] Multimodal acquisition module, used to acquire electroencephalogram signals, virtual / augmented reality interaction behavior data, physiological parameters and gait data of patients, including:

[0079] 1) An electroencephalogram cap, configured to collect the θ / β wave power ratio, P300 event-related potential, and α, β, γ wave band characteristics through electrodes;

[0080] 2) VR / AR device, used to perform virtual supermarket tasks and path navigation tasks, and evaluate executive function and spatial cognitive ability through gesture recognition and voice commands;

[0081] 3) Smart bracelet and piezoelectric insole, configured to continuously monitor the coefficient of variation of step length (CV Step ) and plantar pressure entropy (PDE);

[0082] Data processing module, including:

[0083] 1) Dynamic scale selection engine, initially screened based on a rule engine and optimized by a random forest model, dynamically matching the optimal scale combination according to patient characteristics;

[0084] 2) Multimodal fusion model, which weights and fuses the standardized scores of the scale (Z scale ), electroencephalogram θ / β power ratio (R θ / β), coefficient of variation of gait (CV gait ), standardized value of hippocampal volume (V hippo ) and comprehensive neurotransmitter index (N) through the entropy weight method, and outputs the cognitive impairment risk value;

[0085] 3) Confidence verification module, calculating the internal consistency, multimodal correlation and hospital data support degree of the evaluation result, and the formula is:

[0086] C = 0.35·α + 0.25(|r EEG | + |r gait | + |r N |) + 0.4·I hospital

[0087] α is the internal consistency of the scale, r EEG is the correlation coefficient between the total score of the scale and the θ / β power ratio, r gait is the correlation coefficient between the total score of the scale and the gait parameters, r N is the correlation coefficient between the total score of the scale and the neurotransmitter level, I hospital is the hospital data support degree;

[0088] Edge computing module, deploying a lightweight EEGNet model to a neural processing unit (NPU), processing electroencephalogram signals in real time and uploading key features;

[0089] The human-computer interaction module, including a voice assistant and VR / AR devices, dynamically adjusts the task difficulty and recognizes the user's emotions;

[0090] The output module generates a multi-modal visualization report and realizes dynamic monitoring and early warning through the cloud platform.

[0091] In this embodiment, the multi-modal acquisition module includes:

[0092] 1. EEG cap (Emotiv EPOC+): It collects EEG signals such as the θ / β wave power ratio and P300 event-related potential through highly sensitive electrodes, and simultaneously analyzes the frequency and amplitude changes of α waves, β waves, γ waves, etc. The frequency range of α waves is 8-13Hz, which is usually obvious when the brain is in a relaxed, awake and eyes-closed state, and a decrease in its power may be related to a decline in cognitive function. The frequency of β waves is 13-30Hz, which increases when the brain is excited, thinking nervously or concentrating, and abnormal changes in its power can reflect changes in the excitability of the cerebral cortex. The frequency of γ waves is greater than 30Hz, which is related to high-level cognitive functions such as cognitive processing and memory formation. Through the comprehensive analysis of these EEG signals, information related to brain nerve activities is obtained, providing electrophysiological data support for cognitive function assessment.

[0093] 2. VR / AR devices: The original plan used VR glasses (Oculus Quest 3) to perform virtual supermarket tasks. Now, lightweight interaction improvements are made, and a mobile phone AR scenario (such as "AR virtual supermarket") is added. The Doubao ARCore / ARKit interface (a lightweight AR interface developed based on Android ARCore and iOS ARKit) is called, and the user's gestures and product selection behaviors are captured through the Doubao camera, and the execution function is evaluated in combination with voice commands (such as "Please find the milk"). At the same time, the use scenario of VR glasses is retained. VR glasses perform complex cognitive tasks, such as virtual path navigation, to evaluate spatial cognition and planning capabilities; the mobile phone AR scenario is used for daily simple cognitive assessment and grass-roots applications.

[0094] Specific design of the "VR virtual supermarket task"

[0095] Task process: The subject wears VR equipment (such as Oculus Quest 3) and enters the virtual supermarket scene. The scene contains various common commodities, and the layout is similar to that of a real supermarket. At the start of the task, the system guides the subject to complete tasks such as finding and operating commodities in the virtual supermarket through voice commands (such as "Please find the milk", "Please find the bread and put it in the shopping basket", etc.). The subject interacts through the handle or gestures in the virtual environment, simulating real actions such as walking and picking up commodities. During the process of finding commodities, the system will randomly set some interference factors, such as unclear shelf signs, abnormal commodity placement positions, and the shelf signs rotating randomly by 30°, etc., to increase the task difficulty.

[0096] Evaluation indicators:

[0097] Completion time of the task: The time from the start of the task to when the subject completes all the operations required by the instructions. The shorter the time, the better the executive function and spatial cognitive ability.

[0098] Number of incorrect selections: The number of times the subject selects the wrong product during the process of searching for products. The fewer the number of errors, the higher the cognitive accuracy.

[0099] Path efficiency: Measured by calculating the ratio of the actual walking path length of the subject to the theoretical shortest path length. The closer the ratio is to 1, the stronger the subject's path planning ability in the virtual supermarket.

[0100] 3. Smart bracelet and piezoelectric insole: Real-time monitoring of gait-related data such as the coefficient of variation of step length (CV Step ). The smart bracelet uses sensors such as accelerometers and gyroscopes, combined with specific algorithms, to calculate the step length of each step, and then obtains the coefficient of variation of step length, which is used to measure the degree of change in step length during walking. Multiple piezoelectric sensors are distributed in the piezoelectric insole, which can sense the pressure magnitude in different areas of the sole.

[0101] For the coefficient of variation of step length (CV Step )

[0102]

[0103] Where: σ Step is the standard deviation of the step length, reflecting the volatility of the step length; μ Step is the average value of the step length. CV Step is used to quantify the degree of change in step length over time (in percentage form). The higher the value, the more unstable the gait (such as in the elderly population or patients with neurological diseases).

[0104] For the plantar pressure entropy (PDE):

[0105] Assume that the sole is divided into m regions (m is the number of regions into which the sole is divided). At time t, the pressure value of the jth region is P′ j (t), and the sum of the pressures of all regions is:

[0106]

[0107] Then the probability of the pressure in the jth region:

[0108]

[0109] This formula represents the relationship between P′ j (original value, t) and P total (t), and P j(t) (as a probability) is P′ j (original value, t) divided by P total (t) result; P j (t) is the pressure proportion (i.e., probability) of the jth region at time t;

[0110] The formula of plantar pressure entropy (PDE) is:

[0111]

[0112] The higher the entropy value, the more uneven the pressure distribution (such as flat feet or sports injury risk). By calculating the plantar pressure entropy, the disorder or complexity of the plantar pressure distribution can be reflected, which can help determine the potential impact of cognitive impairment on motor function from the perspective of limb movement.

[0113] 4. Laboratory testing equipment: used to detect neurotransmitter levels, such as acetylcholine, dopamine, γ-aminobutyric acid, etc. Cerebrospinal fluid samples are obtained through lumbar puncture, or peripheral venous blood is collected, and neurotransmitter concentrations are detected using high-performance liquid chromatography (HPLC), enzyme-linked immunosorbent assay (ELISA) and other technologies. These neurotransmitters play a key role in the transmission of neural signals in the brain, and changes in their levels are closely related to the occurrence and development of cognitive dysfunction, providing biochemical data support for the assessment of cognitive impairment.

[0114] In this embodiment, the data processing module includes:

[0115] 1. Dynamic scale selection engine: Based on the rule engine and random forest model to match the optimal scale combination. The rule engine selects the optimal scale combination based on patient characteristics (such as age, years of education, APOE genotype, gait CV Step , cognitive function parameters, neurotransmitter levels, etc.) for preliminary screening. For example, if the education years are ≤6, the initial selection is S = {Graphical MoCA, FAB, RCPM}. These scales are presented in the form of graphics, simple tasks, etc., which are more suitable for patients with lower cultural level and can reduce errors caused by difficulties in understanding text. If APOEε4 is positive, the initial selection is S = {ADAS-Cog, CDR}, because it is related to the risk of Alzheimer's disease, these scales are more targeted for AD assessment. If the patient has poor immediate recall and delayed recall scores in the auditory word learning test, indicating that the memory function is impaired, relevant memory assessment scales such as the Wechsler Memory Scale can be added in a targeted manner. Then, the random forest model is further optimized, and the weight w is obtained through historical data training. j , using the feature matching function f j (X) Calculate the scale suitability score:

[0116]

[0117] This is to ensure the selection of the most suitable assessment scale for the patient, improving the accuracy and pertinence of the assessment. Among them, X is the patient feature matrix, i is the patient index, and j is the feature or scale index.

[0118] Clinical significance of parameter definition

[0119] Patient feature matrix (X):

[0120] It contains multi-dimensional data (demographics, genetics, physiology, cognition, etc.) and is used to comprehensively describe the patient's state. Example: If the patient is positive for APOEε4 + unsteady gait, then AD-specific scales such as ADAS-Cog and CDR are preferentially matched.

[0121] Indices i and j:

[0122] By traversing all patients (i) and scales (j), the optimal combination is dynamically generated. Example: When i = 100 (the 100th patient) and j = 5 (MoCA scale), the system calculates whether this patient is suitable for this scale.

[0123] Multi-modal fusion model: It weights and fuses data such as scales, electroencephalogram, gait, imaging, and neurotransmitter levels. The fusion formula is:

[0124] Risk = 0.25·Z scale +0.35·R θ / β + 0.2·CV gait +0.1·V hippo +0.1·N

[0125] Z scale (Z score ) Scale score, which is obtained by comparing the patient's original score in the scale with the norm data of people of the same age and education level and through standardized conversion, reflecting the patient's relative position in the traditional scale assessment.

[0126] R θ / β is the electroencephalogram R θ / β power ratio, reflecting the state of brain nerve activity.

[0127] The θ wave frequency is 4 - 8 Hz, which is related to cognitive processes such as the brain's relaxation, drowsiness state, and memory encoding. The β wave frequency is 13 - 30 Hz, which is related to excitement and thinking. The change in the ratio of the two can reflect the brain's nerve regulation function.

[0128] CV gait is the gait coefficient of variation CV Step reflecting the patient's gait stability.

[0129] V hippoV is the standardized value of hippocampal volume, which is obtained by segmenting and measuring the hippocampal volume from the patient's brain MRI images, comparing it with the average hippocampal volume of the same age group in the normal population, and through standardization processing. In neurodegenerative diseases such as Alzheimer's disease, hippocampal volume reduction is an important pathological feature.

[0130] N is a comprehensive index of neurotransmitter levels. By performing dimensionality reduction processing such as principal component analysis (PCA) on the concentrations of multiple neurotransmitters (such as acetylcholine, dopamine, gamma-aminobutyric acid, etc.), multiple neurotransmitter variables are converted into a few comprehensive indices, reflecting the overall impact of neurotransmitters on cognitive function.

[0131] 2. The weights are dynamically adjusted by the entropy weight method. The greater the amount of information, the higher the weight. N is the comprehensive index of neurotransmitter levels. Scale Z score (Z scale ) is obtained by comparing the patient's original score in the scale with the norm data of people of the same age and education level, and through standardization conversion, reflecting the patient's relative position in the traditional scale assessment. The electroencephalogram theta / beta power ratio (R θ / β) reflects the state of brain nerve activity. The theta wave frequency is 4 - 8 Hz, which is related to cognitive processes such as the relaxed and drowsy state of the brain and memory encoding. The beta wave is related to excitement and thinking as described above. The change in the ratio of the two can reflect the brain nerve regulation function. (CV gait ) is the coefficient of variation of step length (CV Step ), reflecting the gait stability of the patient. The standardized value of hippocampal volume (V hippo ) is obtained by segmenting and measuring the hippocampal volume from the patient's brain MRI images, comparing it with the average hippocampal volume of the same age group in the normal population, and through standardization processing. In neurodegenerative diseases such as Alzheimer's disease, hippocampal volume reduction is an important pathological feature. The comprehensive index N of neurotransmitter levels is obtained by performing dimensionality reduction processing such as principal component analysis (PCA) on the concentrations of multiple neurotransmitters (such as acetylcholine, dopamine, gamma-aminobutyric acid, etc.), converting multiple neurotransmitter variables into a few comprehensive indices, reflecting the overall impact of neurotransmitters on cognitive function. The specific dimensions after dimensionality reduction by principal component analysis (PCA) of the "comprehensive index N of neurotransmitter levels" (such as "retaining the first 3 principal components"). This model obtains the cognitive impairment risk value through scientific weighted fusion of multi-source data, achieving comprehensive and accurate assessment. The entropy weight method is a method for determining the weights of each index by calculating the entropy value according to the degree of variation of the index data.

[0132] 3. Confidence verification module: Calculate the internal consistency, multimodal consistency, and hospital data support degree. The formula is

[0133] C = 0.35·α + 0.25(|r EEG | + |r gait | + |rN |)+0.4·I hospital

[0134] α is the internal consistency of the scale (Cronbach's α, threshold > 0.7), which measures the reliability and stability of the scale itself by calculating the correlation between the scores of each item in the scale. For example, in a scale containing 10 cognitive test items, if the correlation between the scores of each item is high, it indicates good internal consistency of the scale and can stably measure cognitive function.

[0135] r EEG is the correlation coefficient between the total score of the scale and the θ / β power ratio, reflecting the consistency between the scale evaluation result and the EEG signal;

[0136] r gait is the correlation coefficient between the total score of the scale and gait parameters (such as coefficient of variation of step length, plantar pressure entropy, etc.), reflecting the association between the scale evaluation and limb motor function;

[0137] r N is the correlation coefficient between the total score of the scale and the neurotransmitter level, reflecting the relationship between the scale evaluation and neurotransmitters.

[0138] I hospital is the hospital data support degree (0 or 1, set to 1 if there is hippocampal atrophy, etc.), which provides support for the evaluation result based on the results of hospital imaging examinations (such as MRI showing hippocampal atrophy), genetic tests (such as APOEε4 positive), etc.

[0139] Determine the subsequent operations according to the confidence value C. For example, when C < 0.6, replace the scale combination and supplement hospital examinations to obtain more accurate data; when 0.6 ≤ C < 0.8, increase the VR / AR task difficulty to further examine the patient's cognitive ability; when C ≥ 0.8, generate the final report to ensure the reliability and credibility of the evaluation result.

[0140] In this embodiment, the edge computing module includes: deploying a lightweight EEGNet model to the Doubao NPU (a neural processing unit based on a certain chip architecture, supporting the TensorFlowLite framework) to achieve real-time classification of EEG signals (calculation of θ / β wave power ratio). Only upload the key features to the cloud, reduce the data transmission volume, and improve the processing efficiency. The inference latency at the edge side < 50ms, and the power consumption < 150mW, meeting the requirements of real-time performance and low power consumption. The lightweight EEGNet model is a lightweight neural network model optimized for EEG signal processing, with the characteristics of high computing efficiency and less resource occupancy.

[0141] In this embodiment, the human-computer interaction module includes:

[0142] 1. Voice Interaction and Guidance: Develop a voice assistant (such as "Doubao, start the cognitive test"), and guide users to complete the assessment through multi-round conversations. Dynamically adjust the difficulty of questions, and reduce the task complexity when the error rate is detected to be > 20%. Integrate the BERT model to recognize users' anxiety emotions and trigger comforting voices ("Don't worry, take your time") to enhance the user experience.

[0143] 2. VR / AR Interaction Optimization: Use VR glasses and mobile phone AR scenarios for interaction. In the VR scenario, optimize the immersion of the virtual environment and task design to improve the accuracy of complex cognitive function assessment. In the mobile phone AR scenario ("AR virtual supermarket"), call the Doubao ARCore / ARKit interface to optimize the gesture recognition latency < 200ms to ensure the fluency of interaction. Users complete tasks in the virtual supermarket through simple gesture operations, and the system accurately captures users' behavior data.

[0144] In this embodiment, the output module includes:

[0145] 1. Multimodal Report Generation: Call the Doubao template engine to automatically generate a visual report, such as using a line chart to display the cognitive trend. Support voice broadcast ("Your memory score has increased by 10% compared to last month") to facilitate users' understanding of the assessment results.

[0146] 2. Cloud Platform Dynamic Monitoring and Warning: Support cloud platform dynamic monitoring and warning, real-time upload and analyze patients' multimodal data, timely detect changes in cognitive status and send warning signals, which is convenient for doctors to adjust treatment plans and patients and their families to understand the condition.

[0147] Furthermore, it also includes the following algorithms and layouts:

[0148] 1. Dynamic Scale Selection Algorithm, including

[0149] Input: Patient feature vector X = [age, education years, APOE genotype, gait CV Step , cognitive function parameters, neurotransmitter levels].

[0150] Output: Optimal scale combination S.

[0151] Process: Rule engine preliminary screening, preliminarily screen according to characteristics such as the patient's education years, APOE genotype, cognitive function parameters, neurotransmitter levels, etc. to determine the scale range suitable for the patient. Machine learning optimization, use the random forest model to calculate the scale applicability score, and further optimize the scale selection to ensure that the selected scale can most accurately evaluate the patient's cognitive status.

[0152] 2. Multimodal Fusion Evaluation Model: Comprehensively consider scale Z score, input features such as the electroencephalogram (EEG) θ / β power ratio, gait coefficient of variation, normalized hippocampal volume, and neurotransmitter levels, calculate the risk value through weighted fusion, and achieve a comprehensive assessment of cognitive impairment. This model can fully integrate the information of multi-source data and improve the accuracy and reliability of the assessment.

[0153] 3. Confidence verification mechanism: Quantitatively evaluate the reliability of the assessment results by calculating the internal consistency of the scale, the correlation of multi-modal data, and the support degree of hospital data, and guide the adjustment of subsequent assessment strategies. When the confidence level does not meet the requirements, take corresponding measures in a timely manner to ensure the accuracy and credibility of the assessment results.

[0154] 4. Lightweight EEGNet model: Deployed on the Doubao NPU (a neural processing unit based on a certain chip architecture), it classifies EEG signals in real time, calculates the θ / β wave power ratio, and provides real-time EEG data support for the assessment of cognitive impairment.

[0155] Combined with Figure 2 , the screening and assessment method for elderly cognitive impairment in this embodiment includes the following steps:

[0156] S100. Data acquisition step:

[0157] S110. Collect resting-state and task-state EEG signals through an EEG cap;

[0158] S120. Execute a virtual supermarket task through VR / AR devices, and record the task completion time, the number of wrong selections, and the path efficiency;

[0159] S130. Collect the gait coefficient of variation and plantar pressure entropy through a smart bracelet and piezoelectric insoles;

[0160] S140. Collect neurotransmitter level data and hospital imaging / gene data;

[0161] S200. Dynamic scale selection step:

[0162] S210. Based on the patient's age, education years, APOE genotype, gait parameters, and neurotransmitter levels, initially screen the scale range through a rule engine;

[0163] S220. Use a random forest model to calculate the scale applicability score and optimize the selection of the optimal scale combination;

[0164] S300. Data processing and fusion step:

[0165] S310. Clean, normalize, and standardize the multi-modal data;

[0166] S320. Dynamically weight and fuse the scale scores, EEG features, gait parameters, hippocampal volume, and neurotransmitter indicators through the entropy weight method to calculate the cognitive impairment risk value;

[0167] S400. Confidence verification steps:

[0168] S410. Calculate the internal consistency (Cronbach's α), multimodal data correlation (r EEG 、r gait 、r N ) and hospital data support degree (I hospital ) of the evaluation results to generate a comprehensive confidence value;

[0169] S420. Adjust the evaluation strategy according to the confidence value: if C < 0.6, supplement data and replace the scale; if 0.6 ≤ C < 0.8, increase the task difficulty; if C ≥ 0.8, generate a final report;

[0170] S500. Output and dynamic monitoring steps:

[0171] S510. Generate a multimodal report including a cognitive trend line chart and voice broadcast;

[0172] S520. Real-time monitor data changes through the cloud platform and trigger warning signals.

[0173] In this embodiment, the data collected in the data collection step includes:

[0174] Cognitive function data: Through methods such as the Auditory Verbal Learning Test (AVLT), Digit Cancellation Test, Trail Making Test part B, Boston Naming Test, Clock Drawing Test, etc., collect cognitive function data such as the patient's memory, attention, executive function, language ability, and visuospatial ability.

[0175] Physiological data: Collect physiological data such as EEG activity, brain imaging, neurotransmitter levels, and gait data. Among them, EEG activity data is collected in the resting state and task state through an EEG cap; brain imaging data is obtained through MRI and PET to obtain brain structure and metabolic images; neurotransmitter level data is detected through cerebrospinal fluid or blood samples; gait data is collected by a smart bracelet and piezoelectric insoles. The piezoelectric insoles are made of thin-film piezoelectric sensors, and multiple piezoelectric sensors are evenly distributed on them.

[0176] Other data: Collect the patient's hospital images and genetic data and other physiological parameters to provide a comprehensive data basis for subsequent analysis. During the data collection process, ensure the accuracy and integrity of the data, regularly calibrate and maintain the collection equipment, and at the same time provide detailed guidance and instructions to the patient to ensure that the patient can correctly cooperate to complete various data collection tasks.

[0177] In this embodiment, the dynamic scale selection further includes: the dynamic scale selection engine determines the optimal scale combination based on the patient feature vector through initial screening by the rule engine and optimization by the random forest model. When selecting the scale, individual differences of the patient are fully considered, such as age, educational background, genetic characteristics, cognitive function parameters, neurotransmitter levels, etc., to ensure that the selected scale can accurately evaluate the patient's cognitive function and improve the pertinence and effectiveness of the evaluation.

[0178] In this embodiment, the data processing and fusion further includes: preprocessing the collected data, including operations such as data cleaning and normalization, removing noise and outliers, and unifying the data format and scale. Then, a multi-modal fusion model is used for weighted fusion to calculate the cognitive impairment risk value. At the same time, the electroencephalogram signal is processed in real time through the edge computing module, and only the key features are uploaded to the cloud.

[0179] In this embodiment, human-computer interaction and data collection: The user participates in the cognitive test through the voice assistant and VR / AR devices, and the system collects the user's behavior data and voice data in real time. The voice assistant dynamically adjusts the question difficulty according to the user's performance, and the VR / AR scenario accurately captures the user's gestures and product selection behaviors.

[0180] In this embodiment, the confidence verification further includes: confidence verification and report generation. The confidence verification module calculates the confidence of the evaluation result and decides the subsequent operations according to the confidence value. If the confidence meets the requirements, a multi-modal visualization report is generated; otherwise, the evaluation strategy is adjusted, data is supplemented or the scale is replaced until a reliable evaluation result is achieved.

[0181] In this embodiment, it further includes S600, the verification process of the accuracy of the evaluation algorithm:

[0182] S610, training of the evaluation model

[0183] The cognitive impairment evaluation model is trained using the training set data. The model input is the fused multi-modal data, namely electroencephalogram data, imaging data, physiological parameters, neurotransmitter levels, etc.; the output is the evaluation result of cognitive impairment, which can be specifically divided into different categories such as normal, mild cognitive impairment, dementia, etc. Through learning a large amount of training data, the model continuously optimizes its own parameters to improve the accuracy of cognitive impairment evaluation.

[0184] S620, model verification

[0185] The performance of the trained model is evaluated on the validation set, and a series of key indicators are calculated.

[0186] Accuracy: The calculation formula is This indicator reflects the proportion of the number of correctly predicted samples in the total number of samples by the model, reflecting the overall prediction accuracy of the model.

[0187] Sensitivity: The calculation formula is In the context of cognitive impairment assessment, the number of true positives refers to the number of samples that actually have cognitive impairment and are correctly predicted as diseased by the model, and the number of false negatives refers to the number of samples that actually have cognitive impairment but are wrongly predicted as normal by the model. This indicator reflects the ability of the model to correctly identify samples with cognitive impairment.

[0188] Specificity: The calculation formula is Among them, the number of true negatives refers to the number of samples that are actually normal and are correctly predicted as normal by the model, and the number of false positives refers to the number of samples that are actually normal but are wrongly predicted as having cognitive impairment by the model. This indicator reflects the ability of the model to correctly identify normal samples. A higher specificity can effectively avoid misdiagnosing healthy individuals as cognitive impairment patients, which helps reduce unnecessary subsequent examinations and treatments and lower medical costs in practical applications.

[0189] F1 value (F1 score ):An indicator that comprehensively considers precision and recall rate, and the calculation formula is Its value range is between 0 and 1, and the closer the value is to 1, the better the model performance. The F1 value balances the model's ability to correctly identify patients and healthy individuals, and can more comprehensively evaluate the overall performance of the model, which is of great significance when comparing different models or evaluating the performance of the same model under different conditions.

[0190] S630. Model testing

[0191] Perform a final test on the model that has been verified and optimized on the test set, and calculate indicators such as precision, sensitivity, specificity, and F1 value again to evaluate the generalization ability and accuracy of the model on unknown data. By analyzing the test results, determine whether the model meets the requirements of practical applications. If the model performance does not meet the expectations, the model parameters can be further adjusted, the algorithm can be improved, or the data volume can be increased, etc., and then retrained, verified, and tested until the model performance reaches a satisfactory effect. For example, if it is found that the sensitivity of the model is low in the test set, it may be necessary to increase the number of samples of cognitive impairment patients in the training data, or adjust the feature selection and weight assignment of the model to improve the model's ability to identify cognitive impairment patients. It is also possible to try using ensemble learning methods to fuse multiple models to improve the overall performance of the model.

[0192] S640. Statistical significance test

[0193] Use statistical test methods, such as t-test or chi-square test, to test the consistency between the model evaluation results and the true labels, and determine whether this consistency is statistically significant. For continuous data (such as scale scores, EEG signal eigenvalues, etc.), the t-test can be used to compare whether there is a significant difference in the mean between the model prediction values and the true values. Assume that the scale score predicted by the model is, the true scale score is, and the sample size is, calculate the t-statistic:

[0194]

[0195] , where y 1 and y 2 are the means of the model prediction values and the true values respectively, s p is the pooled standard deviation, n 1 =n 2 =n (assuming the two groups have the same sample size). If the t-test result shows that the p-value is less than the pre-set significance level (such as 0.05), it indicates that the difference between the model evaluation results and the true labels is not caused by random factors, but there is a significant association, further proving the reliability of the model. For categorical data (such as the diagnostic categories of cognitive impairment: normal, mild cognitive impairment, dementia, etc.), the chi-square test can be used to analyze whether the distribution between the model predicted categories and the true categories is consistent. Construct a contingency table and calculate the chi-square statistic

[0196]

[0197] where O ij is the observed frequency (the actual sample size at the intersection of the model predicted category and the true category), E ij is the expected frequency (the theoretical sample size based on the overall distribution assumption), and r and c are the numbers of rows and columns respectively. If the p-value obtained from the chi-square test is less than 0.05, it indicates that there is a significant consistency between the model evaluation results and the true situation in classification, enhancing the credibility of the model.

[0198] S650, ROC Curve Analysis

[0199] Plot the Receiver Operating Characteristic (ROC) curve and calculate the Area Under the Curve (AUC). The ROC curve takes the true positive rate (i.e., sensitivity) as the vertical axis and the false positive rate (1 - specificity) as the horizontal axis. By continuously changing the classification threshold of the model, the true positive rate and false positive rate at different thresholds are obtained, and then the curve is plotted. The calculation formula for AUC is

[0200] AUC = ∫ 0 1 sensitivitiy(x)dx

[0201] The closer its value is to 1, the stronger the classification ability of the model, that is, it can better distinguish samples of different categories. In the assessment of cognitive impairment, a high AUC value means that the model can more accurately distinguish patients from healthy people, providing more effective support for clinical decision-making. For example, if the calculated AUC value of the model is 0.85, it indicates that the model performs well in classification performance and can provide a reliable reference for the diagnosis of cognitive impairment. A higher AUC value means that the model can maintain good discrimination ability at different classification thresholds, and in clinical applications, the diagnostic criteria can be more flexibly adjusted according to actual needs to balance the risks of missed diagnosis and misdiagnosis. At the same time, by comparing the ROC curves and AUC values of different models, the performance advantages and disadvantages of different models can be intuitively evaluated, providing a basis for selecting the optimal model.

[0202] Four actual tests were conducted in the present invention, and the specific situations are as follows:

[0203] Example 1: Assessment of elderly patients in the geriatric department of a general hospital

[0204] Data collection: 100 patients were selected from the geriatric department of a general hospital, and their scores on clinical cognitive assessment scales (MoCA, MMSE), electroencephalogram data (parameters such as power spectral density and coherence of α wave, β wave, θ wave, γ wave, etc. were collected through an Emotiv EPOC+ electroencephalogram cap), brain MRI image data (parameters such as the volume of brain regions such as the hippocampus, amygdala, and frontal lobe and cortical thickness were measured), gait data obtained from smart bracelets and piezoelectric insoles (coefficient of variation of step length, plantar pressure entropy), and concentration data of neurotransmitters (acetylcholine, dopamine, etc.) in serum were collected. At the same time, the true diagnosis results of cognitive impairment were determined based on detailed clinical examinations and expert consultations.

[0205] Model training and verification: The data of 70 patients were used as the training set to train the cognitive impairment assessment model. During the training process, the random forest model described above was used, with the number of trees set to 100, the maximum depth set to 8, the minimum sample split number set to 2, the splitting criterion using Gini impurity, and the feature selection method using random selection of feature subsets (usually when constructing a decision tree in a random forest, a part of the features are randomly selected to find the split points. According to the common practice of general random forests, the size of the feature subset is usually the square root of the total number of features).

[0206] The EEG data was band-pass filtered at 0.5 - 100 Hz, the gray values of MRI images were normalized to the range [0, 1], parameters such as brain region volume were standardized, and outliers in neurotransmitter level data were processed using the three-standard-deviation method. On the validation set (15 patients), the model achieved an accuracy of 82%, a sensitivity of 80%, a specificity of 85%, and an F1 score of 0.81. On the test set (15 patients), the accuracy remained at 80%, demonstrating good generalization ability of the model. The accuracy of single-modal EEG data evaluation was 72%, the accuracy of imaging data evaluation was 75%, the accuracy of neurotransmitter level data evaluation was 68%, and the accuracy increased significantly after multi-modal fusion.

[0207] Confidence verification: The average confidence output by the model was 0.82, higher than the threshold of 0.7. For a small number of patients with low confidence, re-evaluation was carried out to finally ensure the reliability of the evaluation results. Through the confidence calculation formula:

[0208] C = 0.35·α + 0.25(|r EEG | + |r gait | + |r N |) + 0.35·I hospital

[0209] where α is the internal consistency of the scale (Cronbach's α), which was calculated to be 0.8 (greater than the threshold of 0.7), indicating good reliability of the scale itself; r EEG is the correlation coefficient between the total score of the scale and the θ / β power ratio, calculated to be 0.65; r gait is the correlation coefficient between the total score of the scale and gait parameters (coefficient of variation of step length, plantar pressure entropy), calculated to be 0.6; r N is the correlation coefficient between the total score of the scale and neurotransmitter levels, calculated to be 0.55; I hospital is the hospital data support level, determined based on the patient's MRI examination showing hippocampal atrophy, etc. For some patients it is 1, for some it is 0, and the average confidence was calculated to be 0.82.

[0210] Application of the evaluation method: In terms of precise evaluation, by combining the patient's comprehensive physical examination data, detailed medical imaging materials, and multi-scale evaluation results, the model accurately judged the type and degree of the patient's cognitive impairment, providing a key basis for doctors to formulate precise treatment plans. In dynamic evaluation, by long-term monitoring of changes in data such as the patient's behavior, gait, and EEG during VR / AR tasks (such as VR virtual path navigation, AR virtual supermarket tasks), the subtle decline trend of some patients' cognitive function was detected in a timely manner, and the treatment plan was adjusted in advance to delay the progression of the disease.

[0211] For example, a 70-year-old patient was evaluated. The precise evaluation showed that the patient had mild cognitive impairment, mainly manifested as impaired memory and executive function. Dynamic evaluation found that during three consecutive months of monitoring, the patient's reaction time in the VR shopping task gradually increased, and the power of theta waves in the electroencephalogram signal increased. Combining these changes, the doctor timely adjusted the treatment plan, increasing the intensity and frequency of cognitive training, effectively controlling the decline rate of the patient's cognitive function.

[0212] Example 2: Community Screening for Cognitive Impairment

[0213] Data collection: 200 elderly people in a certain community were screened. The Mini-Mental State Examination questionnaire was used for preliminary screening, and then electroencephalogram, VR virtual supermarket task data (through Oculus Quest 3), and gait data were collected using a system. Due to community conditions, imaging data was obtained by cooperating with a nearby hospital to get the information of some elderly people with needs, and neurotransmitter level data was collected by cooperating with a professional testing institution to collect a small amount of blood samples for testing. The true labels were jointly determined by community doctors and hospital experts.

[0214] Model training and verification: After training the model, in the verification stage, the accuracy was 78%, the sensitivity was 75%, and the specificity was 80%. In the unimodal evaluation, the accuracy of VR task data evaluation was 68%, the accuracy of gait data evaluation was 70%, and the accuracy of neurotransmitter level data evaluation was 60%. After multimodal fusion, the accuracy was increased to 78%. In the test stage, most elderly people with cognitive impairment were successfully identified, and the AUC value of the ROC curve was 0.85. During the training process, the data was also preprocessed accordingly. The hyperparameter settings of the random forest model were the same as those in Example 1.

[0215] Confidence verification: Set the confidence threshold at 0.7. Re-evaluate the evaluations with insufficient confidence, adjust the parameters and then evaluate again. Finally, ensure that the evaluation results meet the requirements, effectively screening out 30 elderly people suspected of having cognitive impairment, providing a basis for subsequent referral and further diagnosis. For evaluations that do not meet the confidence requirements, check whether there are abnormalities in the data, such as interference noise in electroencephalogram data, incorrect records of gait data, etc. Correct or re-collect the abnormal data, and at the same time adjust the hyperparameters of the random forest model, such as trying to change the number of trees, the maximum depth, etc. Retrain the model and then evaluate again until the confidence meets the requirements. The number of trees is 100, the maximum depth is 8, the minimum sample split number is 2, the splitting criterion uses Gini impurity, and the feature selection method uses randomly selected feature subsets (the size of the feature subset is the square root of the total number of features).

[0216] Application of assessment methods: In community screening, precise assessment combines limited physical examination data, imaging and neurotransmitter data obtained through cooperation, and the results of scale screening to preliminarily judge the cognitive impairment risk of the elderly. Dynamic assessment continuously monitors the performance of the screened elderly in daily simple VR / AR cognitive tasks (such as the AR virtual supermarket item-finding task), gait changes, etc., to timely detect changes in potential cognitive impairment risks, providing dynamic information support for community health management and facilitating early intervention. For example, during the screening of a 65-year-old elderly person, precise assessment found that their cognitive function was in a critical state. Through dynamic assessment, it was found that the error rate of this elderly person gradually increased and the coefficient of variation of step length also increased during subsequent AR cognitive tasks. Based on these changes, the community doctor promptly referred the elderly person to a higher-level hospital for further examination, enabling the elderly person to receive timely diagnosis and treatment.

[0217] Example 3: Follow-up assessment of patients in a rehabilitation institution

[0218] Data collection: There are 50 patients in the recovery period of cognitive impairment in a rehabilitation institution. Regularly collect their cognitive assessment scale data (scales customized for the recovery stage), electroencephalogram data during rehabilitation training, VR / AR data simulating daily activities (such as using AR to simulate a home scene to complete daily tasks), and gait data during gait rehabilitation training. At the same time, regularly detect the neurotransmitter levels in serum or cerebrospinal fluid. Rehabilitation experts determine the true cognitive recovery situation based on the rehabilitation progress and reexamination results of the patients.

[0219] Model training and verification: Train a model to evaluate the cognitive improvement of patients during the rehabilitation process. In the validation set, the accuracy of the model's judgment on cognitive improvement is 85%, the sensitivity is 88%, and the specificity is 82%. Multimodal fusion assessment can accurately capture the subtle cognitive changes of patients during the rehabilitation process, while single-modal data is relatively weak in reflecting the comprehensiveness of the rehabilitation effect. During the training process, according to the data characteristics of patients in the recovery period, appropriately adjust the data preprocessing parameters. For example, when filtering electroencephalogram data, the cut-off frequency of the filter can be adjusted according to the specific interference frequencies that may occur during rehabilitation training. The hyperparameters of the random forest model are still optimized through grid search combined with cross-validation. The hyperparameters are optimized through grid search combined with cross-validation. Suppose that after optimization, the number of trees is determined to be 100, the maximum depth is 8, the minimum number of samples for splitting is 2, the splitting criterion uses Gini impurity, and the feature selection method uses randomly selected feature subsets (the size of the feature subset is the square root of the total number of features).

[0220] Confidence Verification: The confidence calculation results show that the confidence levels of most evaluation results are above 0.8. For individual evaluations with low confidence, re-analyze in combination with rehabilitation training records and doctor observations to ensure the accuracy of the evaluation of the patient's rehabilitation progress and provide a scientific basis for adjusting the rehabilitation plan. For example, for a patient with low confidence, by checking the rehabilitation training records, it is found that there is a contradiction between the EEG signals and behavioral performances during the execution of specific tasks. Further analysis reveals that the abnormal EEG data is caused by interference in the training environment. After re-collecting and processing the data, the confidence level of the evaluation result of this patient is increased to above 0.8.

[0221] Application of Evaluation Methods: In rehabilitation institutions, accurate evaluation relies on various physiological index changes during the patient's rehabilitation, the scores of performance evaluation scales in rehabilitation training, and imaging reexamination data to accurately judge the degree of the patient's cognitive recovery. Dynamic evaluation monitors the patient's performance in VR / AR tasks during rehabilitation training, gait improvement, and real-time changes in EEG signals in real time, and compares with historical data to adjust the rehabilitation training plan in a timely manner and improve the rehabilitation effect. For example, a patient who has been hospitalized for a long time due to stroke for rehabilitation shows that at the initial stage of rehabilitation, accurate evaluation shows that his cognitive function is in a moderately damaged state, mainly manifested as attention and executive function disorders. During the rehabilitation process, dynamic evaluation finds that when the patient performs VR-simulated daily tasks, the time to complete the tasks gradually shortens, the error rate decreases, and at the same time, the power of β waves related to attention in the EEG signals increases. According to these dynamic evaluation results, the rehabilitation therapist timely adjusts the rehabilitation training plan, increases the difficulty and complexity of the tasks, and further promotes the recovery of the patient's cognitive function. After several months of rehabilitation training, a re-accurate evaluation shows that the patient's cognitive function has been significantly improved and reaches the level of mild damage.

[0222] Example 4: Health Management of the Elderly in Nursing Homes

[0223] Data Collection: Health monitoring is carried out on 150 elderly people in a nursing home, collecting daily cognitive observation records, EEG and gait data continuously monitored by wearable devices, VR / AR cognitive test data regularly carried out (such as regular AR cognitive mini-games), and regularly collecting blood samples to detect neurotransmitter levels. The nursing home doctors and external cooperative medical institutions jointly determine the true labels of the cognitive states of the elderly.

[0224] Model training and verification: After model training, in the verification stage, the accuracy of the classification of the cognitive state of the elderly was 80%, the sensitivity was 78%, and the specificity was 83%. Multimodal fusion evaluation has obvious advantages in long-term health management and can timely detect the changing trend of the cognitive state of the elderly. The accuracy of single-modal data such as long-term EEG data evaluation is 70%, the accuracy of long-term gait data evaluation is 73%, and the accuracy of long-term neurotransmitter level data evaluation is 65%, which is increased to 80% after fusion. During the training process, considering the long-term accumulation and change characteristics of the data of the elderly in the nursing home, the data is segmented and analyzed, and the hyperparameters of the random forest model are optimized and adjusted according to the data of different time periods. According to the data of different time periods, the optimization and adjustment are performed. It is assumed that the number of adjusted trees is 100, the maximum depth is 8, the minimum number of sample splits is 2, the splitting criterion uses Gini impurity, and the feature selection method uses random selection of feature subsets (the size of the feature subset is the square root of the total number of features).

[0225] Confidence verification: By setting the confidence threshold to 0.75, the assessments that did not meet the confidence standard were reviewed and re-evaluated after supplementing the data. The system helped the nursing home to promptly discover the deterioration of the cognitive status of the five elderly people and take intervention measures in advance. For assessments that did not meet the confidence standard, additional data on the elderly’s recent living habits, sleep quality, etc. were collected and included in the model analysis. At the same time, the model was retrained and evaluated to improve the reliability of the assessment results.

[0226] Application of assessment methods: During the precise assessment, the nursing home's regular physical examination data, the neurotransmitter levels of blood tests, the results of imaging examinations (if any), and the scores of various cognitive assessment scales are integrated to establish an accurate cognitive health profile for each elderly person. Dynamic assessment uses long-term monitored VR / AR cognitive test data, EEG and gait data, and uses time series models to analyze the trend of cognitive function changes. If it is found that an elderly person's reaction time is significantly prolonged and the coefficient of variation of step length increases in several consecutive VR / AR tests, combined with the power changes in specific frequency bands in the EEG signal, it is judged that his cognitive function may decline. The nursing home arranges more personalized cognitive training activities for the elderly and strengthens nursing attention. Specifically, when the nursing home conducts regular assessments for a 75-year-old man, the precise assessment shows that his cognitive function is within the normal range, but the dynamic assessment finds that in the past six months, his reaction time in VR cognitive tests has gradually increased from an average of 30 seconds to 45 seconds, the coefficient of variation of step length has also increased from 0.06 to 0.09, and the power of EEG alpha waves has decreased. Based on these data, the nursing home developed a special cognitive training plan for the elderly, including daily memory training games and simple logical thinking exercises, while the nursing staff increased the frequency of interaction with the elderly. After a period of intervention, the re-evaluation found that the elderly's cognitive function indicators tended to stabilize and did not deteriorate further.

[0227] It should be noted that unless otherwise specified, the technical terms or scientific terms used in this application should have the ordinary meanings understood by those skilled in the art to which this application pertains.

[0228] As mentioned above, the above are only the preferred specific embodiments of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in this application should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A screening and evaluation system for elderly cognitive impairment, characterized in that: include: Multimodal acquisition module, used to collect EEG signals, virtual reality (VR) / augmented reality (AR) interaction behavior data, gait parameters and neurotransmitter level data; Data processing module, including: Dynamic scale selection engine, based on rule engine initial screening and random forest model optimization, dynamically matches the optimal scale combination according to patient characteristics; Multimodal fusion model, weighted fusion scale standardized score (Z scale ), EEG θ / β power ratio (R θ / β), gait coefficient of variation (CV gait ), hippocampal volume normalized value (V hippo ) and neurotransmitter comprehensive index (N), output the risk value of cognitive impairment, and dynamically adjust the weight through the entropy weight method; The confidence verification module calculates the internal consistency, multimodal correlation and hospital data support of the evaluation results. The formula is: C=0.35·α+0.25(|r EEG |+|r gait |+|r N |)+0.4·I hopspital α is the internal consistency of the scale, r EEG is the correlation coefficient between the total score of the scale and the θ / β power ratio, r gait is the correlation coefficient between the total score of the scale and the gait parameters, r N is the correlation coefficient between the total score of the scale and the level of neurotransmitters, I hospital Support for hospital data; The human-computer interaction module, including voice assistants and VR / AR devices, dynamically adjusts task difficulty and recognizes user emotions.

2. The screening evaluation system according to claim 1, characterized in that: The multimodal acquisition module also includes: An EEG cap, configured to collect theta / beta wave power ratio, P300 event-related potential, and alpha, beta, and gamma wave frequency band characteristics through electrodes; Virtual reality (VR) / augmented reality (AR) devices are used to perform virtual supermarket tasks and path navigation tasks, and to assess executive function and spatial cognition abilities through gesture recognition and voice commands; Smart bracelet and piezoelectric insole, configured to monitor the coefficient of variation (CV) of stride length in real time Step ) and plantar pressure entropy (PDE); Laboratory testing equipment used to measure neurotransmitter levels.

3. The screening evaluation system according to claim 2, characterized in that: The smart bracelet is worn on the wrist of the user, and the piezoelectric insole is placed in the user's shoe; the smart bracelet calculates the length of each step of the user, and then obtains the coefficient of variation of the step length, which is used to measure the degree of change of the step length during walking; There are multiple piezoelectric sensors distributed in the piezoelectric insole, which can sense the pressure in different areas of the sole of the foot; For the coefficient of variation of step length (CV Step ) Where: Step is the standard deviation of the step length, reflecting the volatility of the step length; μ Step is the average value of the step length; CV Step Used to quantify the degree of temporal variability in stride length, with higher values ​​indicating more unstable gait; For plantar pressure entropy (PDE): Assume that the sole of the foot is divided into m regions (m is the number of regions divided into the sole of the foot). At time t, the pressure value of the jth region is P′ j (t), the sum of all regional pressures is: Then the probability of pressure in the jth region is: This formula represents P′ j (original value, t) and P total (t), P j (t) is the pressure proportion of the jth region at time t; The formula of plantar pressure entropy (PDE) is: The higher the entropy value, the more uneven the pressure distribution. By calculating the plantar pressure entropy, the degree of disorder or complexity of the plantar pressure distribution can be reflected, which can help determine the potential impact of cognitive impairment on motor function from the perspective of limb movement.

4. The screening evaluation system according to claim 2, characterized in that: The data processing module also includes: Dynamic scale selection engine: Based on the rule engine and random forest model matching the optimal scale combination, the rule engine performs preliminary screening based on patient characteristics; if the education years are ≤6, the initial selection is S = {graphical MoCA, FAB, RCPM}; if APOEε4 is positive, the initial selection is S = {ADAS-Cog, CDR}; if the patient's immediate recall and delayed recall scores in the auditory word learning test are poor, indicating that the memory function is impaired, relevant memory assessment scales such as the Wechsler Memory Scale are added in a targeted manner; then, the random forest model is further optimized, and the weight w is obtained through historical data training j , using the feature matching function f j (X) Calculate the scale suitability score: This ensures that the most suitable assessment scale for the patient is selected, improving the accuracy and pertinence of the assessment; where X is the patient feature matrix, i is the patient index, and j is the feature or scale index.

5. The screening evaluation system according to claim 2, characterized in that: The multimodal fusion model also includes: weighted fusion scale, EEG, gait, imaging and neurotransmitter level data, and the fusion formula is: Risk=0.25·Z scale +0.35·R θ / β+0.2·CV gait +0.1·V hippo +0.1·N Z scale The scale score is obtained by comparing the patient's original score in the scale with the normative data of people of the same age group and education level, and then undergoing standardized transformation to reflect the patient's relative position in the traditional scale assessment; R θ / β is EEG R θ / β power ratio, reflecting the state of brain nerve activity; The frequency of theta waves is 4-8Hz, which is related to cognitive processes such as relaxation, sleepiness, and memory encoding. The frequency of beta waves is 13-30Hz, which is related to excitement and thinking. The change in the ratio of the two can reflect the brain's neural regulation function. CV gait CV is the coefficient of variation of gait Step Reflects the patient's gait stability; V hippo The standardized value of hippocampal volume is obtained by segmenting the patient's brain MRI image, measuring the hippocampal volume, and comparing it with the mean hippocampal volume of the same age group in the normal population after standardization; N is a comprehensive indicator of neurotransmitter levels.

6. The screening evaluation system according to claim 2, characterized in that: The confidence verification module also includes: calculating internal consistency, multimodal consistency and hospital data support, the formula is: C=0.35·α+0.25(|r EEG |+|r gait |+|r N |)+0.4·I hospital α is the internal consistency of the scale (Cronbach's α, threshold > 0.7), which measures the reliability and stability of the scale itself by calculating the correlation between the scores of each item in the scale; r EEG is the correlation coefficient between the total score of the scale and the θ / β power ratio, reflecting the consistency between the scale evaluation results and the EEG signals; r gait is the correlation coefficient between the total score of the scale and the gait parameters, reflecting the association between the scale assessment and limb motor function; r N is the correlation coefficient between the total score of the scale and the neurotransmitter level, reflecting the relationship between the scale assessment and neurotransmitters; I hospital To provide support for the evaluation results based on the hospital's imaging and genetic testing results; Subsequent operations are determined based on the confidence value C. For example, when C<0.6, the scale combination is replaced and additional hospital examinations are conducted to obtain more accurate data. When 0.6≤C<0.8, the difficulty of the VR / AR task is increased to further examine the patient's cognitive ability. When C≥0.8, a final report is generated to ensure the reliability and credibility of the evaluation results.

7. A screening and assessment method for elderly cognitive impairment, characterized in that: The steps include: S100, data collection steps: S110, collecting resting state and task state EEG signals through an EEG cap; S120, executing a virtual supermarket task through a VR / AR device, and recording the task completion time, the number of wrong choices, and the path efficiency; S130, collecting step length variation coefficient and plantar pressure entropy through smart bracelet and piezoelectric insole; S140, collect neurotransmitter level data and hospital imaging / genetic data; S200, Dynamic scale selection steps: S210, based on the patient's age, years of education, APOE genotype, gait parameters and neurotransmitter levels, the scale range was initially screened through the rule engine; S220, use the random forest model to calculate the scale applicability score and optimize the selection of the optimal scale combination; S300, data processing and fusion steps: S310, cleaning, normalizing and standardizing the multimodal data; S320. Calculate the risk value of cognitive impairment by dynamically weighting the fusion scale score, EEG characteristics, gait parameters, hippocampal volume and neurotransmitter indexes using the entropy weight method; S400, confidence verification steps: S410, calculate the internal consistency (Cronbach's α) and multimodal data correlation (r EEG 、r gait 、r N ) and hospital data support (I hospital ), generate a comprehensive confidence value; S420, adjust the evaluation strategy according to the confidence value: if C<0.6, supplement the data and change the scale; if 0.6≤C<0.8, increase the task difficulty; if C≥0.8, generate a final report; S500, output and dynamic monitoring steps: S510, generating a multimodal report including a cognitive trend line graph and a voice broadcast; S520. Monitor data changes in real time through the cloud platform and trigger early warning signals.

8. The screening and evaluation method according to claim 7, characterized in that: In the dynamic scale selection step, the hyperparameters of the random forest model are set as: number of trees = 100, maximum depth = 8, minimum number of sample splits = 2, and the splitting criterion is Gini impurity.

9. The screening and evaluation method according to claim 7 or 8, characterized in that: It also includes S600 and the verification process for evaluating the accuracy of the algorithm: S610, Evaluation Model Training Use the training set data to train the cognitive impairment assessment model; S620, Model Verification Evaluate the performance of the trained model on the validation set and calculate the accuracy, sensitivity, and specificity; S630, Model Test Perform a final test on the verified and optimized model on the test set, and recalculate indicators such as accuracy, sensitivity, specificity, and F1 value to evaluate the generalization ability and accuracy of the model on unknown data; by analyzing the test results, determine whether the model meets the needs of actual applications; If the model performance does not meet expectations, you can further adjust the model parameters, improve the algorithm or increase the amount of data, and retrain, verify and test until the model performance reaches satisfactory results; S640, Statistical significance test Using statistical test methods, the consistency between the model evaluation results and the true labels is tested to determine whether this consistency is statistically significant; For continuous data, the t-test can be used to compare whether there is a significant difference in the mean between the model predicted value and the true value; Assume that the scale score predicted by the model is, the true scale score is, and the sample size is, calculate the t statistic: Where y1 and y2 are the means of the model prediction value and the true value respectively, s p is the combined standard deviation, n1 and n2 are the sample sizes; if the t-test result shows that the p-value is less than the pre-set significance level, it means that the difference between the model evaluation result and the true label is not caused by random factors, but there is a significant correlation; For categorical data, the chi-square test can be used to analyze whether the distribution between the model predicted category and the actual category is consistent; construct a contingency table and calculate the chi-square statistic: Among them, ij is the observation frequency, E ij is the expected frequency, r and c are the number of rows and columns respectively; S650, ROC curve analysis The receiver operating characteristic (ROC) curve was drawn and the area under the curve (AUC) was calculated. The ROC curve uses the true positive rate as the ordinate and the false positive rate as the abscissa. By continuously changing the classification threshold of the model, the true positive rate and false positive rate under different thresholds were obtained, and then the curve was drawn. The calculation formula of AUC is: The closer its value is to 1, the stronger the classification ability of the model is, that is, it can better distinguish samples of different categories.

10. The screening and evaluation method according to claim 9, characterized in that: The S620 also includes: Accuracy is calculated as follows: This indicator reflects the proportion of samples predicted correctly by the model in the total number of samples, reflecting the overall prediction accuracy of the model; Sensitivity, calculated as: In the context of cognitive impairment assessment, the number of true positive cases refers to the number of samples that actually suffer from cognitive impairment and are correctly predicted by the model as having the disease, and the number of false negative cases refers to the number of samples that actually suffer from cognitive impairment but are incorrectly predicted by the model as normal. This indicator reflects the ability of the model to correctly identify samples with cognitive impairment; Specificity is calculated as follows: Among them, the number of true negative cases refers to the number of samples that are actually normal and correctly predicted as normal by the model, and the number of false positive cases refers to the number of samples that are actually normal but are incorrectly predicted as having cognitive impairment by the model; this indicator reflects the ability of the model to correctly identify normal samples; F1 value: An indicator that comprehensively considers accuracy and recall. The calculation formula is: The value of F1 ranges from 0 to 1. The closer the value is to 1, the better the model performance is.

Images