An early cognitive impairment prediction system

Abstract

This application relates to the field of smart healthcare technology, specifically disclosing an early cognitive impairment prediction system, comprising: determining the optimal single-frequency detection point from multi-frequency EEG data using an optimal single-frequency point determination module, and extracting bioimpedance features at that frequency point using a data acquisition module; screening key risk factors using a risk factor acquisition module; constructing sample sets with multiple input modes using a sample set construction module, and training multiple basic screening models using the sample set construction module respectively; integrating the basic models using an early cognitive impairment prediction model construction module based on a voting mechanism to form a composite screening model, and using this composite screening model to assess the risk level of the test subjects. This invention integrates electrophysiological characteristics and clinical information, achieves data dimensionality reduction through an attention mechanism, and improves model robustness through voting integration, ultimately achieving efficient, sensitive, and interpretable early cognitive impairment assisted screening.

Description

Technical Field

[0001] This application relates to the field of smart healthcare technology, and in particular to an early cognitive impairment prediction system. Background Technology

[0002] Mild cognitive impairment (MCI) is a common disease among the elderly, caused by various brain lesions (including cerebral hemorrhage, cerebral infarction, cerebral atrophy, and brain tumors). Timely diagnosis of MCI can significantly improve patients' quality of life and effectively delay the progression to dementia. Currently, clinical identification of MCI mainly relies on neuropsychological scales and imaging examinations. While neuropsychological scales are simple and easy to administer, they suffer from high subjectivity and insufficient standardization. Imaging examinations (such as MRI), while providing objective evidence, have limitations such as expensive equipment, complex procedures, and long processing times, making them difficult to widely apply in community screening and large-scale population surveys.

[0003] In existing technologies, for example, CN121415816A proposes a mild cognitive impairment screening system. This system simultaneously collects speech and drawing data, extracts speech envelope and handwriting velocity, determines a low-frequency collaborative window based on power spectrum, calculates the speech-writing lag index, and extracts listening and motion disturbances to eliminate non-cognitive interference. Finally, it uses the debiased residuals for Bayesian risk classification. However, this approach has the following drawbacks: the overall process is complex and computationally burdensome, making real-time deployment in primary healthcare or home settings difficult; it heavily relies on high-precision synchronous acquisition equipment, imposing stringent requirements on hardware and operating environments; it assumes a linear relationship between speech-writing lag and hearing and motion disturbances, potentially oversimplifying the complex interaction between cognitive and physiological factors; and the method relies on multi-step manual feature engineering, failing to fully leverage the end-to-end advantages of deep learning, and the rationality of the feature design lacks sufficient validation. Summary of the Invention

[0004] The purpose of this invention is to provide an early cognitive impairment prediction system that partially solves or alleviates the above-mentioned shortcomings of the prior art, and can achieve efficient, sensitive and interpretable early cognitive impairment assisted screening.

[0005] To solve the aforementioned technical problems, the present invention specifically adopts the following technical solution: A first aspect of the present invention is to provide a method for predicting early cognitive impairment, comprising the following steps: Brain bioimpedance data above / below the occipital bone were acquired from a cranial electromagnetic field detection device for both MCI and healthy subjects at multiple preset frequencies. The brain bioimpedance data included amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope. A first attention module is constructed by training the first attention module with brain bioimpedance data above / below the occipital bone collected by the cranial electromagnetic field detection device at multiple preset frequencies and the corresponding subject classification labels, so as to obtain a first attention module that can output attention weight sequences at each frequency. The optimal single frequency point for model construction is determined according to the attention weight sequences at each frequency, and the brain bioimpedance data above / below the occipital bone collected at the optimal single frequency point is obtained based on the optimal single frequency point. Collect multiple basic clinical indicators and physiological parameters related to the subjects and standardize them to obtain standardized parameters; calculate the standardized mean difference of each standardized parameter between the MCI subject group and the healthy subject group, and rank the standardized parameters according to the standardized mean difference to screen at least one risk factor; Based on the brain bioimpedance data above / below the occipital bone collected by the brain electromagnetic field detection device at the optimal single frequency point and the screened risk factors, a training sample set is constructed. The training sample set is divided using a preset pattern to obtain six initial sample sets, including: The upper half of the input sample set includes: amplitude, phase, amplitude difference, and phase difference in the occipital region; The lower half of the input sample set includes: suboccipital amplitude, phase, amplitude difference, and phase difference; The dual-semi-input sample set includes: occipital bone amplitude, occipital bone phase, occipital bone amplitude difference, and occipital bone phase difference; The complete input sample set includes: occipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The complete input sample set includes: suboccipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The dual complete input sample set includes: occipital vertical amplitude, occipital vertical phase, occipital vertical amplitude difference, occipital vertical phase difference, amplitude slope, and phase slope. For each initial sample set, a multi-layer feedforward neural network was used for training and cross-validation to obtain multiple basic screening models; Based on the aforementioned multiple basic screening models and voting mechanisms, an early cognitive impairment prediction model is constructed. The brain bioimpedance data of the subject above / below the occipital bone at the optimal single frequency point are obtained from the cranial electromagnetic field detection device and input into the early cognitive impairment prediction model to predict the subject, so as to obtain the prediction result, which is the risk level of the subject for early cognitive impairment.

[0006] The risk level output by the predictive model in this application is only a reference indicator for medical staff and is not a final diagnostic result. For example, if the subject is predicted to have a high risk level of early cognitive impairment, the medical staff will, based on this reference indicator, have the subject undergo further diagnostic procedures such as computed tomography (CT) and magnetic resonance imaging (MRI), or use the Montreal Cognitive Assessment (MoCA) to complete a standardized diagnosis.

[0007] Furthermore, the first attention module is a one-dimensional attention module; Furthermore, during the training process, the first attention module is trained by maximizing the matching degree between the weighted feature sequence of the one-dimensional attention module and the subject's classification label, so as to obtain the first attention module that can output attention weight sequences of each frequency.

[0008] Furthermore, the construction of an early cognitive impairment prediction model based on the aforementioned multiple basic screening models and voting mechanism includes: Map the output results of each basic screening model to the corresponding risk level; If all basic screening models output the same risk level, then output that risk level. If any basic screening model outputs a high risk and no other basic screening model outputs a low risk, then output a high risk. If any basic screening model outputs low risk and no other basic screening model outputs high risk, then output low risk. If a basic screening model exists that outputs both high-risk and low-risk values, then the output will be medium-risk.

[0009] Furthermore, the collected basic clinical indicators and physiological parameters are divided into binary indicators and continuous indicators; wherein, the binary indicators are encoded in binary using 0 or 1, and the continuous indicators are converted to the 0-1 range using a normalization method.

[0010] Furthermore, it also includes the following steps: If the output of a basic screening model differs from the output of other models by more than a preset anomaly threshold, then the basic screening model will be temporarily retrained or temporarily excluded from this vote.

[0011] Furthermore, the risk factors include one or more of the following: whether or not one is married, whether or not one has coronary heart disease, whether or not one has hypertension, and whether or not one has hyperglycemia.

[0012] Furthermore, it also includes the following steps: For subjects of different genders, separate initial sample sets were constructed for males and females, and gender-related basic screening models were trained separately.

[0013] Furthermore, the prediction of the test subjects based on the early cognitive impairment prediction model also includes: Receive the personal information of the test subject, which includes at least age information; The age information is compared with a preset age threshold; If the age information exceeds a preset age threshold, an enhanced screening process is initiated. The enhanced screening process includes forcibly collecting information on whether the subject has a spouse, and inputting this information as a risk factor into an early cognitive impairment prediction model for prediction.

[0014] Furthermore, the optimal single frequency point is 50kHz.

[0015] Secondly, this application also discloses an early cognitive impairment prediction system, the system comprising: The data acquisition module is configured to acquire brain bioimpedance data above / below the occipital bone from the cranial electromagnetic field detection device at multiple preset frequencies for MCI / healthy subjects. The brain bioimpedance data includes amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope. The optimal single-frequency point determination module is configured to construct the first attention module. It uses brain bioimpedance data above / below the occipital bone collected by the cranial electromagnetic field detection device at multiple preset frequencies and the corresponding subject classification labels to train the first attention module to obtain a first attention module that can output attention weight sequences at each frequency. The optimal single-frequency point for model construction is determined based on the attention weight sequences at each frequency, and the brain bioimpedance data above / below the occipital bone collected at the optimal single-frequency point is obtained based on the optimal single-frequency point. The risk factor acquisition module is configured to collect multiple basic clinical indicators and physiological parameters related to the subjects, and perform standardization processing to obtain standardized parameters; calculate the standardized mean difference of each standardized parameter between the MCI subject group and the healthy subject group, and sort the standardized parameters according to the standardized mean difference to screen at least one risk factor; The sample set construction module is configured to construct a training sample set based on the brain bioimpedance data above / below the occipital bone collected by the cranial electromagnetic field detection device at the optimal single frequency point and the screened risk factors. The sample set partitioning module is configured to partition the training sample set using a preset mode to obtain six initial sample sets, including: The upper half of the input sample set includes: amplitude, phase, amplitude difference, and phase difference in the occipital region; The lower half of the input sample set includes: suboccipital amplitude, phase, amplitude difference, and phase difference; The dual-semi-input sample set includes: occipital bone amplitude, occipital bone phase, occipital bone amplitude difference, and occipital bone phase difference; The complete input sample set includes: occipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The complete input sample set includes: suboccipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The dual complete input sample set includes: occipital vertical amplitude, occipital vertical phase, occipital vertical amplitude difference, occipital vertical phase difference, amplitude slope, and phase slope. The basic screening model training module is configured to train and cross-validate multiple basic screening models using a multi-layer feedforward neural network for each initial sample set. An early cognitive impairment prediction model construction module is configured to construct an early cognitive impairment prediction model based on the multiple basic screening models and the voting mechanism. The prediction module is configured to acquire brain bioimpedance data above / below the occipital bone of the subject under the optimal single frequency point from the cranial electromagnetic field detection device, and input the data into the early cognitive impairment prediction model to predict the subject and obtain a prediction result, which is the risk level of the subject for early cognitive impairment.

[0016] Furthermore, the first attention module is a one-dimensional attention module; and during the training process, the first attention module is trained by maximizing the matching degree between the weighted feature sequence of the one-dimensional attention module and the subject's classification label, so as to obtain the first attention module that can output attention weight sequences of each frequency.

[0017] Furthermore, the early cognitive impairment prediction model construction module is specifically configured to: map the output results of each basic screening model to the corresponding risk level; if the output risk level of all basic screening models is consistent, then output that risk level; if any basic screening model outputs a high risk and no other basic screening model outputs a low risk, then output a high risk; if any basic screening model outputs a low risk and no other basic screening model outputs a high risk, then output a low risk; if there are basic screening models that output both high and low risk, then output a medium risk.

[0018] Furthermore, the risk factor acquisition module is also configured to divide the collected basic clinical indicators and physiological parameters into binary indicators and continuous indicators; wherein, the binary indicators are encoded in binary using 0 or 1, and the continuous indicators are converted to the 0-1 range using a normalization method.

[0019] Furthermore, the early cognitive impairment prediction model construction module is also configured to: if the output of a certain basic screening model differs from the output of other models by more than a preset abnormal threshold, then temporarily retrain the basic screening model or temporarily exclude it from the current vote.

[0020] Furthermore, the risk factors include one or more of the following: whether or not one is married, whether or not one has coronary heart disease, whether or not one has hypertension, and whether or not one has hyperglycemia.

[0021] Furthermore, the basic screening model training module is configured to construct male-specific initial sample sets and female-specific initial sample sets for subjects of different genders, and train gender-related basic screening models respectively.

[0022] Furthermore, the prediction module is specifically configured to receive the personal information of the subject to be tested, which includes at least age information; compare the age information with a preset age threshold; if the age information exceeds the preset age threshold, initiate an enhanced screening process, which includes forcibly collecting information on whether the subject to be tested has a spouse, and inputting the information on whether the subject has a spouse as a risk factor into the early cognitive impairment prediction model for prediction.

[0023] Furthermore, the optimal single frequency point is 50kHz.

[0024] Furthermore, the optimal single-frequency point determination module is specifically configured as follows: A feature vector is constructed from the raw brain bioimpedance data collected above and / or below the occipital bone at each of multiple preset frequency points. ,in Let be the feature dimension at this frequency point, and arrange the feature vectors of all frequency points in frequency order to form a two-dimensional feature matrix. , as input to the attention module; For the input two-dimensional feature matrix Through three weight matrices respectively , and Perform a linear transformation to generate the query matrix. Key matrix Sum matrix : ;in, For the dimensions of query and key, For the dimension of the value, ; Then through the query matrix AND key matrix The dot product is used to calculate the similarity score between each pair of frequency points, and after scaling and softmax normalization, the attention weight matrix is ​​obtained. : ; And the overall attention weight sequence for each frequency point The attention weight matrix is ​​averaged in the value dimension: Among them, the first Line number Column elements Indicates the first The frequency point is paired with the first Attention weights for each frequency point; Then, the attention weight matrix Acting on the value matrix The weighted feature sequence is obtained. : ; Weighted feature sequence Flattened into a one-dimensional vector, it is sequentially input into a fully connected layer and a softmax output layer to obtain the predicted probability that the subject belongs to the MCI category. ; The cross-entropy loss function is used to measure the difference between the predicted probability and the true label. Differences between them: ; With the goal of minimizing cross-entropy loss, the learnable parameters in the attention module are backpropagated and iteratively updated.

[0025] Beneficial technical effects: This application significantly improves feature discrimination efficiency while reducing data acquisition complexity through a data-driven frequency optimization mechanism; it enhances the robustness and clinical interpretability of the model in practical applications through a multimodal information fusion strategy that integrates electrophysiological features and clinical risk factors; and it effectively integrates discrimination information from different feature perspectives through an ensemble learning-based voting mechanism, significantly improving the stability and reliability of screening results while balancing sensitivity and specificity. This application supports efficient and low-cost rapid initial screening of large-scale populations and can be seamlessly embedded into a hierarchical medical system, providing reliable auxiliary initial screening tools for community medical institutions. Through modular design, it achieves decoupling between the algorithm and hardware, facilitating flexible deployment and iterative optimization in different clinical scenarios, and providing a practical technical path for promoting the early detection and intervention of cognitive impairment. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale. Obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0027] Figure 1 This is a flowchart of a method for predicting early cognitive impairment according to this application.

[0028] Figure 2 This is a schematic diagram of MCI risk factor input in one embodiment of this application.

[0029] Figure 3 This is a schematic diagram of the prediction result interface obtained by using the MCI screening model to predict the results of a healthy subject A in one embodiment of this application.

[0030] Figure 4 This is a schematic diagram of the prediction result interface obtained by using the MCI screening model to predict the results of a healthy subject B in one embodiment of this application.

[0031] Figure 5 This is a schematic diagram of the prediction result interface obtained by using the MCI screening model to predict the prediction of a healthy subject C in one embodiment of this application.

[0032] Figure 6 This is a schematic diagram of the prediction result interface obtained by using the MCI screening model to predict the prediction of healthy subject D in one embodiment of this application.

[0033] Figure 7This is a schematic diagram of the prediction result interface obtained by using the MCI screening model to predict the results of a healthy subject E in one embodiment of this application.

[0034] Figure 8 This is a schematic diagram of the prediction result interface obtained by using the MCI screening model to predict the prediction of healthy subject F in one embodiment of this application.

[0035] Figure 9 This is a schematic diagram of the module structure of an early cognitive impairment prediction system in one embodiment of this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] In this document, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" may be used interchangeably.

[0038] In this document, the terms "upper," "lower," "inner," "outer," "front," "rear," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the present invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0039] In this document, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0040] In this document, "and / or" includes any and all combinations of one or more of the listed related items.

[0041] In this article, "multiple" means two or more, that is, it includes two, three, four, five, etc.

[0042] Figure 1 The flowchart of an early cognitive impairment prediction method (which can also be called a brain bioimpedance data processing method) of this application is shown below. Figure 1 The method specifically includes the following steps: S1. Obtain brain bioimpedance data above / below the occipital bone from MCI subjects and healthy subjects at multiple preset frequencies using a cranial electromagnetic field detection device. The brain bioimpedance data includes amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope.

[0043] In one embodiment of the present invention, a raw dataset for training a mild cognitive impairment (MCI) screening model is first constructed. The raw dataset includes multi-source information from a large number of subjects, divided into MCI patient cases and cognitively normal healthy controls. All subjects were diagnosed by professional physicians using the Montreal Cognitive Assessment (MoCA) and assigned specific category labels, namely "MCI" or "healthy," as the gold standard for subsequent supervised learning.

[0044] Furthermore, brain bioimpedance data were collected for each subject. Specifically, a four-electrode method was used to perform frequency sweep measurements within the frequency range of 10Hz to 100kHz, focusing on acquiring impedance response data in the region above the occipital bone (defined as the upper half region) and the region below the occipital bone (defined as the lower half region). The bioimpedance data included (impedance) amplitude, (impedance) phase, amplitude difference, phase difference, amplitude slope, and phase slope.

[0045] In some embodiments, the brain electromagnetic field detection device may be an existing non-invasive dynamic monitor for brain edema.

[0046] S2. Construct a first attention module. Train the first attention module using brain bioimpedance data above / below the occipital bone collected by the cranial electromagnetic field detection device at multiple preset frequencies and the corresponding subject classification labels to obtain a first attention module that can output attention weight sequences at each frequency. Determine the optimal single frequency point for model construction based on the attention weight sequences at each frequency. Obtain brain bioimpedance data above / below the occipital bone collected by the cranial electromagnetic field detection device at the optimal single frequency point.

[0047] In some embodiments, S2 includes: Assume there is a total A feature vector is formed by collecting raw brain bioimpedance data (including amplitude, phase, etc.) above and / or below the occipital bone at each preset frequency point (e.g., 10kHz, 20kHz, ..., 100kHz). ,in This represents the feature dimension at that frequency point. Arrange the feature vectors of all frequency points in frequency order to form a two-dimensional feature matrix. , which serves as the input to the attention module.

[0048] For the input feature matrix Using three learnable weight matrices respectively , and Perform a linear transformation to generate the query matrix. Key matrix Sum matrix : ; in For the dimensions of query and key, The dimension of the value is usually taken as .

[0049] Then, by querying the matrix AND key matrix The dot product is used to calculate the similarity score between each pair of frequency points, and after scaling and softmax normalization, the attention weight matrix is ​​obtained. : ; Among them, matrix The first in Line number Column elements Indicates the first The frequency point is paired with the first Attention weights for each frequency point.

[0050] To obtain the overall attention weight sequence for each frequency point The attention weight matrix is ​​averaged in the value dimension: ; The weight sequence It reflects the relative contribution of each frequency point to the classification result and serves as the basis for determining the optimal frequency point after training is completed.

[0051] Obtaining the attention weight matrix Then, the attention weight matrix Acting on the value matrix The weighted feature sequence is obtained. : ; That is, the output feature of each frequency point is the weighted sum of the value vectors of all frequency points at that frequency point.

[0052] Weighted feature sequence Flattened into a one-dimensional vector, it is sequentially input into a fully connected layer and a softmax output layer to obtain the predicted probability that the subject belongs to the MCI category. The cross-entropy loss function is used to measure the relationship between the predicted probability and the true label. Differences between (MCI=1, Health=0): ; To minimize the cross-entropy loss, the Adam optimizer is used to optimize all learnable parameters (including) in the attention module. Backpropagation and iterative updates are performed on the parameters of the fully connected layers. The training process uses batch input (batchsize=32), the learning rate is initially set to 0.001, and an early stopping mechanism is set. Training is terminated when the validation set loss no longer decreases for 10 consecutive rounds.

[0053] After training, the attention weight sequence for each frequency point is extracted from the attention module. The average attention weight of multiple subjects on the validation set was calculated, and the frequency point with the highest average weight was selected as the optimal single frequency point. In multiple independent training and five-fold cross-validation tests, the 50kHz frequency point had the highest average attention weight (0.43), which was significantly better than other frequency points (all others were below 0.3), so it was determined as the optimal single frequency point.

[0054] In some embodiments, to select the most discriminative single-frequency detection points from multi-band EEG data, this method constructs a first attention module (a one-dimensional self-attention module) to learn the importance distribution of different frequency points for the early cognitive impairment (MCI) classification task. This module takes multi-band EEG data as input, automatically learns frequency weights by optimizing the classification target, and ultimately determines the optimal single-frequency point for model construction, achieving data dimensionality reduction and feature focusing, thereby improving model efficiency and interpretability.

[0055] The first attention module is a one-dimensional self-attention structure. Its input consists of brain bioimpedance data collected above and / or below the occipital bone at multiple preset frequencies (10kHz, 20kHz, ..., 100kHz), including amplitude and phase information at each frequency. The multi-frequency data for each subject is arranged in frequency order as a one-dimensional feature sequence. The module's training objective is to maximize the match between its output weighted feature sequence and the subject's classification label (MCI patient or healthy individual). During training, the module calculates the dependencies between frequency points through a self-attention mechanism and assigns an attention weight to each frequency point, ultimately outputting an attention weight sequence that reflects the relative contribution of each frequency to the classification result.

[0056] As mentioned earlier, the self-attention mechanism in this embodiment measures the similarity between each pair of frequency point feature vectors by calculating the dot product of the query matrix and the key matrix. This similarity is the dependency between frequency points. The higher the similarity, the more related the EEG physiological response patterns of the two frequency points are. Subsequently, the above similarity scores are scaled and normalized using Softmax to obtain the attention weight matrix. Each element in the matrix represents the attention weight of one frequency point to another. Then, mean pooling is performed on the value dimension of this matrix to obtain the overall attention weight sequence corresponding to each frequency point. The weight values ​​in this sequence directly reflect the relative contribution of the corresponding frequency point to the MCI classification result.

[0057] After training, the attention weight sequence output by the first attention module was extracted and statistically analyzed. The frequency point with the highest weight was determined as the optimal single-frequency point for model construction. Experimental results show that in multiple independent training and cross-validation processes, the attention weight corresponding to the 50kHz frequency point was consistently significantly higher than that of other frequency points, demonstrating good stability and discriminative significance. Therefore, the system selected 50kHz as the preferred single-frequency point for subsequent model construction and data acquisition.

[0058] After determining the optimal single-frequency point, the system automatically filters and extracts brain bioimpedance data collected above and / or below the occipital bone at that frequency point (50kHz), including its amplitude, phase, and related derived parameters (such as amplitude difference, phase difference, slope, etc.). This data will serve as the core input features for the subsequent construction and operation of the MCI screening model, thereby significantly reducing the complexity of data acquisition and computational burden while ensuring model performance, achieving an efficient transition from multi-frequency exploration to single-frequency optimization.

[0059] S3. Collect multiple basic clinical indicators and physiological parameters related to the subjects and standardize them to obtain standardized parameters; calculate the standardized mean difference of each standardized parameter between the MCI subject group and the healthy subject group, and sort the standardized parameters according to the standardized mean difference to screen at least one risk factor.

[0060] To enhance the comprehensive discriminative ability of early cognitive impairment screening models, this method systematically integrates the subjects' basic clinical indicators and physiological parameters as auxiliary features. First, raw data containing multiple types of information is collected, including binary indicators such as whether the subject has coronary heart disease, hypertension, hyperglycemia, has a spouse, and various medical histories and lifestyle habits, as well as continuous indicators such as heart rate, systolic blood pressure, diastolic blood pressure, and blood glucose levels. Next, this raw data is standardized. For binary indicators, a uniform binary encoding is used, with "yes" or "absent" encoded as 1 and "no" or "not present" encoded as 0. For continuous indicators, a minimum binary encoding is used. The maximum value normalization method linearly maps each index to the [0, 1] interval to eliminate the dimensional differences between different indices, thereby forming a standardized set of parameters that can be used for model training.

[0061] To further identify the most valuable features for distinguishing early cognitive impairment from healthy status among the standardized parameters, this method introduces the standardized mean difference as a basis for ranking parameter importance. For each standardized parameter, its mean is calculated in both the group of subjects diagnosed with early cognitive impairment and the healthy group. The difference between the two means is then calculated and divided by the pooled standard deviation calculated for both groups, thus obtaining the standardized mean difference of the parameter. The standardized mean difference reflects the magnitude of the difference in the distribution of the parameter between the two groups. The larger the absolute value, the stronger the correlation between the parameter and the state of early cognitive impairment, and the higher its potential as a risk factor.

[0062] In some embodiments, based on the calculated standardized mean difference, all standardized parameters are sorted from highest to lowest absolute value, resulting in a parameter sequence arranged in descending order of discriminant importance. Based on this sorting result, appropriate thresholds can be set or the top few parameters can be selected to identify the risk factors most indicative of early cognitive impairment. In practical applications, the key risk factors identified by this method mainly include whether or not one has a spouse, whether or not one has coronary heart disease, whether or not one has hypertension, and whether or not one has hyperglycemia. These factors not only show a statistically significant association with early cognitive impairment but also conform to clinical epidemiological understanding and possess clear medical interpretability.

[0063] Ultimately, the selected risk factors, along with key features such as brain bioimpedance data, will be used as input to the model to construct a screening model capable of making joint judgments by comprehensively utilizing multi-source information. This screening method based on statistical difference quantification avoids the bias that may be introduced by subjective feature selection and improves the robustness and interpretability of the model in practical applications, providing a reliable feature engineering foundation for efficient and accurate screening of early cognitive impairment.

[0064] S4. Based on the brain bioimpedance data above / below the occipital bone collected by the cranial electromagnetic field detection device at the optimal single frequency point and the screened risk factors, a training sample set is constructed.

[0065] Based on the optimal single-frequency point determined in the preceding steps and the selected key risk factors, this method proceeds to the training sample set construction stage. The training sample set is the foundation of model learning, and its construction quality directly affects the performance and generalization ability of the final screening model. This step aims to systematically integrate electrophysiological data and clinical risk information to construct a standardized sample set with high discriminative power and a clear structure, providing sufficient and balanced data support for the subsequent training of the deep learning model.

[0066] Specifically, firstly, based on the established optimal single frequency of 50kHz, brain bioimpedance data of all subjects (MCI subjects and healthy subjects) at this frequency, collected by a cranial electromagnetic field detection device, are extracted from the original dataset, focusing on the area above and / or below the occipital bone. This data includes, but is not limited to: raw amplitude and phase measurements at this frequency, amplitude difference and slope calculated based on amplitude, and phase difference and slope calculated based on phase. The bioimpedance data for each subject will be organized into a structured feature vector, encompassing multi-dimensional information about the EEG response at this frequency.

[0067] Simultaneously, key risk factors identified during screening, including whether a subject is married, whether they have coronary heart disease, hypertension, and hyperglycemia, are correlated and integrated with the aforementioned bioimpedance data. Each subject's risk factor status is coded based on their clinical records or questionnaire results: for binary factors, 0 (no / no) or 1 (yes / no) are used; for continuous factors (such as blood pressure, blood glucose levels, etc., which, if included, require normalization), their standardized values ​​are used. These risk factors serve as auxiliary features, together with the bioimpedance features, to form a complete comprehensive feature vector.

[0068] Furthermore, each subject sample must have a clear category label. This label is determined based on the diagnostic results of the MOCA scale: if a subject is diagnosed with MCI, they are labeled as a positive sample (label 1); if they are a healthy subject, they are labeled as a negative sample (label 0). The label information serves as the training objective for supervised learning, ensuring that the model can learn the mapping relationship from the comprehensive feature vector to the MCI state.

[0069] The comprehensive feature vectors and corresponding labels of all the subjects are organized in a uniform format to form a complete training sample set. Before being put into use, this sample set usually needs to undergo necessary sample balancing (such as oversampling of minority class samples or undersampling of majority class samples) to avoid model training bias caused by class imbalance. The completed training sample set combines fine-grained information of electrophysiological features with discriminative information of clinical risk factors.

[0070] To optimize the discrimination performance of the screening model and fully consider the differences in physiological characteristics and risk factor distribution among subjects of different genders, this invention further constructs separate screening models for men and women.

[0071] In the model construction process, the overall sample is first divided according to gender, forming a subset of male subjects and a subset of female subjects. For the male subset, brain bioimpedance features collected at the optimal single frequency point (e.g., 50kHz) are extracted, and risk factors with higher significance in the male population (such as whether or not they have a spouse, history of coronary heart disease, etc.) are integrated. A multi-layer feedforward neural network is then used for independent training to obtain a male-specific screening model. Similarly, based on the female subset and its corresponding combination of features and risk factors, a female-specific screening model is trained. The two models maintain the same structure but exhibit gender specificity in network weights and decision thresholds, thereby more accurately capturing subtle differences in early cognitive impairment manifestations between male and female subjects and improving the sensitivity and specificity of the overall screening system.

[0072] S5. The training sample set is divided using multiple modes to obtain six initial sample sets, including: The upper half of the input sample set includes: amplitude, phase, amplitude difference, and phase difference in the occipital region; The lower half of the input sample set includes: suboccipital amplitude, phase, amplitude difference, and phase difference; The dual-semi-input sample set includes: occipital bone amplitude, occipital bone phase, occipital bone amplitude difference, and occipital bone phase difference; The complete input sample set includes: occipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The complete input sample set includes: suboccipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The dual complete input sample set includes: occipital vertical amplitude, occipital vertical phase, occipital vertical amplitude difference, occipital vertical phase difference, amplitude slope, and phase slope. The six initial sample sets correspond to six preset input modes, including: dual-half input mode (D0), upper half input mode (D1), lower half input mode (D2), dual-complete input mode (S0), upper complete input mode (S1), and lower complete input mode (S2). "Dual" indicates that bioimpedance features from both the upper and lower regions of the occipital bone are used simultaneously; "upper" or "lower" indicates that features from only the upper or lower regions of the occipital bone are used; "half" indicates that the input features do not include a slope term (i.e., only amplitude, phase, amplitude difference, and phase difference, a total of four dimensions); and "complete" indicates that the input features include all six bioimpedance features (i.e., amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope).

[0073] In some embodiments, when constructing the input for the screening model, all input patterns are fused with the aforementioned clinical risk factor vector through feature concatenation. Specifically, for any input pattern (e.g., upper half input pattern, double full input pattern, etc.), its original feature vector contains brain bioimpedance data extracted under that pattern, such as amplitude, phase, difference, and slope. Simultaneously, the subject's clinical risk factor vector (including binary factor vector and continuous factor vector) is extracted in parallel. Before inputting into the model, the impedance feature vector and the clinical risk factor vector are directly concatenated along the feature dimension to form a unified, high-dimensional composite feature vector, which serves as the complete input to the neural network input layer. This fusion method retains all the information of the original signal features while introducing risk priors with clinical discriminative value, enabling the model to collaboratively learn the correlation patterns between physiological signals and health risks, thereby improving the comprehensive discriminative ability for early cognitive impairment.

[0074] S6. For each initial sample set, a multi-layer feedforward neural network is used for training and cross-validation to obtain multiple basic screening models.

[0075] For each of the aforementioned initial sample sets, this method performs model training and validation. The initial sample sets are divided into six sample sets based on the data acquisition location (above the occipital bone, below the occipital bone, or a combination thereof) and feature type (whether it includes amplitude slope and phase slope). Each sample set contains feature vectors extracted from the corresponding data pattern and category labels determined based on the MOCA scale method, representing a specific data view or feature combination, aiming to characterize the subject's electroencephalological state from different perspectives.

[0076] To fully utilize the information contained in each initial sample set, this method employs a multi-layer feedforward neural network as the basic learning architecture, independently modeling each initial sample set. The neural network is uniformly configured as a three-layer feedforward architecture comprising an input layer, at least one hidden layer, and an output layer. The hidden layer neurons use the hyperbolic tangent function as the activation function, and the output layer uses the logistic function to map the network output to a probability value belonging to early cognitive impairment. For different input sample sets, the dimension of the neural network's input layer automatically adapts according to the length of the corresponding feature vector, and the number of hidden layer neurons is optimized within a preset range; for example, the number of neurons in the first hidden layer is between 8 and 16, and the number of neurons in the second hidden layer is between 4 and 8, aiming to achieve a balance between model expressive power and generalization ability.

[0077] In some embodiments, each basic screening model employs a three-layer MLNN structure, including an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is dynamically determined based on the selected input mode; for example, for the half-mode, the input dimension is 4 (bioimpedance) + 4 (risk factors) = 8; for the full mode, the input dimension is 6 + 4 = 10. The hidden layer uses the hyperbolic tangent function (tanh) as the activation function, and its number of nodes is optimized within a preset range through cross-validation. The output layer contains a single neuron, uses the logistic activation function, and has an output value between 0 and 1, representing the probability that the subject is diagnosed with MCI.

[0078] To objectively evaluate the performance of each model and avoid overfitting during training, this method employs a cross-validation strategy to perform multiple partitions and training iterations on each initial sample set. Specifically, a random 7:3 partitioning strategy is used ten times, randomly dividing each initial sample set into a training subset (70%) and a validation subset (30%), repeated ten times. After each partition, the training subset is used to perform supervised training on a multi-layer feedforward neural network, minimizing the prediction error through backpropagation. Binary cross-entropy is used as the loss function, combined with an adaptive learning rate optimizer (such as Adam) for parameter updates. An early stopping mechanism is implemented during training, with the convergence criterion being that the loss on the validation set no longer significantly decreases. After training, the model performance is tested using an independent validation subset, recording key metrics such as sensitivity and specificity. Through ten repeated experiments, ten intermediate basic screening models trained under different data partitions are obtained for each initial sample set.

[0079] Finally, for each initial sample set, the performance metrics of the intermediate models obtained in ten cross-validations are averaged. This average performance is used as the final performance estimate of the basic screening model for that data pattern. The model parameters that best perform on the validation set from the ten training iterations are selected as the basic screening model corresponding to that pattern. Thus, six basic screening models will be generated for each of the six initial sample sets. Each model focuses on learning a discrimination pattern from its specific data perspective, laying the foundation for subsequently building a more robust composite screening model through an ensemble strategy.

[0080] S7. Based on the aforementioned multiple basic screening models and voting mechanism, construct an early cognitive impairment prediction model.

[0081] After obtaining multiple basic screening models, this method constructs a composite screening model to integrate the discriminative advantages of different basic models, thereby improving the robustness and stability of the final screening decision. The basic screening models are trained based on different data collection sites and feature combinations. A single model may have limitations in some cases, while the composite model can achieve a more comprehensive and robust identification of early cognitive impairment through collaborative decision-making.

[0082] Specifically, the composite screening model is constructed using a voting mechanism as its core integration strategy. This mechanism employs six basic screening models as voting members. Each model, when predicting the risk level of a subject for early cognitive impairment, outputs its risk level. The risk level is determined based on the probability value output by the model: if the probability value is higher than a preset first threshold, it is classified as high risk; if the probability value is lower than a preset second threshold (which is lower than the first threshold), it is classified as low risk; and if it falls between the two, it is classified as medium risk. The voting mechanism makes a comprehensive decision based on the outputs of the six basic models, according to predefined rules.

[0083] In some embodiments, the first threshold and the second threshold are determined by the following method: First, a validation dataset is constructed for threshold optimization. This validation dataset is independent of the training set and includes both diagnosed MCI subjects and healthy subjects, with the true classification labels of all subjects known. The size of the validation set can be determined based on the overall sample size, typically accounting for 20% to 30% of the total sample size.

[0084] Secondly, the search range and step size for the thresholds are determined. Since the probability values ​​output by the basic screening model are between 0 and 1, the search range for both the first threshold (high-risk threshold) and the second threshold (low-risk threshold) is set to 0 to 1. To balance search accuracy and computational efficiency, a search step size of 0.1 is used, meaning the search is iterated over 11 candidate threshold points: 0.0, 0.1, 0.2, ..., 1.0. It should be noted that the first threshold is used to determine high risk; that is, when the model output probability is greater than or equal to the first threshold, it is considered high risk. The second threshold is used to determine low risk; that is, when the model output probability is less than or equal to the second threshold, it is considered low risk. When the probability value is between the second and first thresholds, it is considered medium risk. Therefore, the constraint that the first threshold must always be greater than the second threshold must be satisfied during the iteration process.

[0085] Then, the objective function is selected for optimization. This application uses the Youden index as the evaluation index for threshold optimization. The Youden index is calculated as sensitivity plus specificity minus 1, where sensitivity represents the proportion of subjects with actual MCI who are correctly identified as high-risk or medium-risk, and specificity represents the proportion of subjects with actual health conditions who are correctly identified as low-risk. The Youden index ranges from -1 to 1; a higher value indicates a stronger overall discriminative ability of the screening model.

[0086] Next, threshold pair traversal and optimization are performed. For each threshold combination that satisfies the condition that the first threshold is greater than the second threshold, the corresponding Youden index is calculated on the validation set. Specifically, for each subject in the validation set, the probability values ​​output by the six basic screening models are first obtained. Then, based on the current threshold combination, the probability values ​​of each model are mapped to high-risk, medium-risk, or low-risk levels. A voting mechanism is then used to obtain the final risk level of the composite model. Finally, the output of the composite model is compared with the subject's true label to calculate the sensitivity and specificity under that threshold combination, thus obtaining the Youden index. After traversing all possible threshold combinations, the threshold pair that maximizes the Youden index is selected as the optimal threshold.

[0087] In one specific embodiment of this application, the validation set is threshold optimized according to the above method, and the final optimal threshold pair is: the first threshold equals 0.7, and the second threshold equals 0.3. Under this threshold combination, the Youden index of the composite screening model reaches its maximum value on the validation set.

[0088] It should be noted that the above thresholds (0.7 for high risk and 0.3 for low risk) were obtained through iterative optimization based on the experimental data used in this application. Alternatively, risk thresholds can be set in a conventional manner, such as setting the first threshold to any value between 0.6 and 0.8, and the second threshold to any value between 0.2 and 0.4, or flexibly adjusted according to actual clinical needs and the trade-off between screening sensitivity and specificity.

[0089] In some embodiments, the voting rules are as follows: 1. If all base models output the same risk level, the composite model will directly output that level. 2. If at least one basic model outputs high risk and no basic model outputs low risk, then the composite model outputs high risk. 3. If at least one basic model outputs low risk and no basic model outputs high risk, then the composite model outputs low risk. 4. If the basic model outputs both high and low risk, and the composite model outputs medium risk, further clinical evaluation is recommended.

[0090] This voting mechanism emphasizes both consistency in risk warnings (such as issuing a decisive warning when most models indicate a high risk) and a prudent middle ground when there are significant differences between models, thereby improving the ability to detect potential early cognitive impairment cases without excessively increasing false positives.

[0091] In some embodiments, to further enhance the reliability and fault tolerance of the composite screening model in the integrated decision-making process, this invention introduces an anomaly detection and adaptive processing mechanism for the output results of the basic screening models. This mechanism continuously monitors the risk level or predicted probability value output by each basic screening model during the voting process and compares it in real time with the mode or average of the outputs of the remaining basic models. If a significant deviation is detected between the output value of a basic model and the overall trend, and this deviation exceeds a preset anomaly threshold, the system automatically determines that the current output of that model is abnormal and immediately initiates the corresponding dynamic processing procedure. For example, the abnormal model may undergo rapid incremental learning based on recent samples, updating its parameters in a limited number of steps to adapt to the current data distribution; or the model may be temporarily isolated from the valid models in this vote, preventing it from participating in the final risk integration calculation for the current subjects. Regardless of the processing method used, the system records the anomaly event and processing log for subsequent model performance analysis and optimization. This mechanism can effectively identify and mitigate integration decision biases caused by short-term performance fluctuations of individual models, data noise interference, or feature distribution shifts, thereby significantly enhancing the stability and robustness of the composite screening system in complex application environments.

[0092] In some embodiments, the anomaly detection threshold is determined based on a statistical distribution on a validation set. First, a validation dataset, independent of the training set, is constructed. This validation set includes diagnosed MCI subjects and healthy subjects whose true classification labels are known. The size of the validation set typically accounts for 20% to 30% of the total sample size. Then, all K basic screening models are run on the validation set. In this application, K equals 6, meaning there are six basic models, denoted as Model 1, Model 2, Model 3, Model 4, Model 5, and Model 6. For each subject in the validation set, the six probability values ​​output by all six models for that subject are obtained. Each probability value is between 0 and 1, representing the likelihood that the model considers the subject to have MCI. Then, the deviation is calculated for each model: taking Model 1 as an example, the arithmetic mean of the output probabilities of the five models (Models 2 to 6) is calculated first, then the output probability of Model 1 is subtracted from this mean and the absolute value is taken. The resulting value is the deviation of Model 1 for this subject. Similarly, for Model 2, the mean of the output probabilities of the five models (Models 1, 3, 4, 5, and 6) is calculated, then the output probability of Model 2 is subtracted from this mean and the absolute value is taken. This process continues until the deviation of Model 6 is calculated. In this way, each model obtains a deviation value each time. The smaller the value, the more consistent the model's judgment is with the overall judgment of the other five models; the larger the value, the more isolated the model's judgment is, and the more likely it is to be in an abnormal state. The deviations of all subjects and all six models are summarized and statistically analyzed to obtain a sequence containing N times 6 deviation values. The mean μ and standard deviation σ of this sequence are calculated, and the abnormal judgment threshold T is set to μ plus k times σ, where k is an adjustable parameter used to control the strictness of abnormal judgment. The larger the value of k, the higher the threshold and the stricter the anomaly detection criteria; only models with extremely high deviations will be identified as anomalous. Conversely, the smaller the value of k, the lower the threshold and the more lenient the anomaly detection criteria, allowing more models to be flagged as anomalous. In one specific embodiment of this application, k is set to 2; in another embodiment with higher stability requirements, k is set to 3. Those skilled in the art can also flexibly adjust the value of k within the range of 1.5 to 3.5 according to the trade-off between fault tolerance and sensitivity required in practical applications. When the current deviation of a base model under the test subjects exceeds the threshold T, the system determines that the current output of the model is in an anomalous state and immediately initiates the corresponding dynamic processing procedure, such as temporarily retraining the model or temporarily excluding it from the current vote; if the current deviation does not exceed the threshold T, the model output is determined to be normal, and its result participates normally in subsequent voting integration.

[0093] The final composite screening model, while maintaining the discrimination expertise of each basic model, effectively smooths out the predictive fluctuations that may be caused by data noise or feature limitations of a single model through integrated voting, thereby achieving higher clinical applicability and decision reliability overall. In other words, steps S1-S7 above are actually a method for constructing a predictive model for predicting the risk level of early cognitive impairment; that is, this predictive model is used to predict the specific risk level of a subject being a patient with early cognitive impairment.

[0094] S8. Obtain brain bioimpedance data above / below the occipital bone of the subject under the optimal single frequency point from the cranial electromagnetic field detection device, and input it into the early cognitive impairment prediction model to predict the subject to obtain a prediction result, wherein the prediction result is the risk level of the subject to have early cognitive impairment.

[0095] After the composite screening model is constructed, it can be applied to real-world scenarios for automated screening of early cognitive impairment in test subjects. This process follows standardized operating procedures to ensure the consistency and reliability of the screening results. Specifically, when screening test subjects, relevant data must first be collected according to standards consistent with the training phase. This includes acquiring brain bioimpedance data above and / or below the occipital bone at the optimal single frequency (50kHz), and collecting information on key risk factors such as spousal status, coronary heart disease, hypertension, and hyperglycemia. This data must undergo the same preprocessing and standardization steps as the training samples to form a comprehensive feature vector conforming to the model input format.

[0096] Subsequently, the preprocessed and standardized feature vectors are input into the six basic screening models included in the composite screening model. Each basic model performs forward computation independently and outputs a predicted probability value for the subject to have early cognitive impairment. Based on the preset thresholds of each basic model, these probability values ​​are further converted into high-risk, medium-risk, or low-risk classifications. This step realizes the mapping from continuous probability values ​​to discrete risk categories, preparing for subsequent voting integration.

[0097] Next, the voting mechanism in the composite screening model is activated. This mechanism aggregates the risk level assessments from the six basic models and makes an integrated decision based on predefined voting rules. The voting rules aim to comprehensively consider the judgments of each basic model. If all models are consistent, the judgment is adopted directly; if there is controversy regarding the coexistence of high and low risk, a medium risk is output to alert clinical attention; if only a single risk indication is given (e.g., only high risk or no low risk), then high or low risk is output accordingly. Through this mechanism, the composite model can effectively mitigate abnormal outputs from individual models while leveraging the strengths of each basic model, thereby forming a more robust and reliable final screening opinion.

[0098] In some embodiments, the final visualization result is as follows Figures 2 to 8 As shown, Figure 2 This is a schematic diagram of the risk factor input interface. Users can enter corresponding risk factors on this interface according to the actual situation of different subjects. Figures 3 to 7 This is a schematic diagram illustrating the visualization results on the device for six different subjects. Figure 3 For illustration, the top left corner of the figure shows a multi-model risk assessment radar chart, displaying the independent judgment results of six basic sub-models (upper half model, lower half model, upper complete model, lower complete model, dual half model, and dual complete model) using a hexagonal coordinate system. Each vertex corresponds to the model name, and the radial distance represents the risk level (low-medium-high), determined by a voting mechanism as high risk. The bottom left corner shows the perturbation coefficient (PC) indicator, one of the standard output parameters of existing bioimpedance detection equipment. The perturbation coefficient can be used as an auxiliary examination indicator to help identify signal abnormalities caused by factors such as body movement, respiration, emotional fluctuations, or external interference. It can form a double screening with the risk level determination, effectively preventing false alarms caused by temporary interference and improving the reliability of screening results. It is shown in a red high-level bar chart in the figure, indicating the presence of significant physiological perturbation in the current state. The right side of the figure shows the curve of the perturbation coefficient changing over time.

[0099] Specifically, see Figure 3 As can be seen, the outputs of the lower half model, the upper complete model, the lower complete model, the double half model, and the double complete model are all high-risk, while the output of the upper half model is medium-risk. Therefore, the final prediction result is high-risk.

[0100] See Figure 4 As can be seen, the upper half model, upper complete model, lower complete model, double half model, and double complete model all output high risk, while the lower half model and lower complete model output medium risk. Therefore, the final prediction result is high risk.

[0101] See Figure 5 As can be seen, the upper half model, lower half model, double complete model, and lower complete model all output low risk, while the upper complete model and double half model output medium risk. Therefore, the final prediction result is low risk.

[0102] See Figure 6 As can be seen, the upper half model, lower half model, upper complete model, and lower complete model all output low risk, while the double complete model and double half model output medium risk. Therefore, the final prediction result is low risk.

[0103] See Figure 7As can be seen, the lower half model and the lower complete model both output medium risk, the upper complete model outputs low risk, while the double complete model, the double half model, and the upper half model all output high risk. Therefore, the final prediction result is medium risk.

[0104] See Figure 8 As can be seen, the output of the bi-half model is high risk, the output of the lower half model is medium risk, and the rest are low risk. Therefore, the final prediction result is medium risk.

[0105] In some embodiments, before screening the subject, the system can obtain the subject's personal information, including at least age information. The system compares this age information with a preset age threshold, typically set based on the age group with a significantly increased risk of cognitive impairment in epidemiological data; for example, the age threshold might be set to 60 years old. If the subject's age exceeds the preset threshold, an enhanced screening process is automatically triggered. In this process, the system will require the collection or supplementation of the subject's spousal status information, clarifying whether they currently have a spouse—a binary variable—and input this information as an important risk factor, along with other physiological characteristics and clinical indicators, into the early cognitive impairment prediction model for comprehensive risk assessment and classification decisions. Through this age-triggered information enhancement mechanism, the system can automatically improve the risk profile for high-risk age groups, particularly strengthening the weight of sociological factors in screening, thereby improving the targeting and predictive accuracy of screening for specific populations.

[0106] In some embodiments, the system outputs the screening results for the subject, typically presented as a structured report. The report includes at least the determined risk level (high, medium, or low), and may optionally include output details, confidence indices, or brief explanations of each base model. If the output is high or medium risk, the system may further recommend more detailed neuropsychological evaluations (such as a full MOCA scale test) or clinical diagnoses such as MRI. This screening process automates the entire chain from data acquisition to result output, ensuring screening efficiency while providing quantifiable and interpretable decision-making support for the initial identification of early cognitive impairment.

[0107] Table 1. Comparison of traditional scale methods and the screening method of this invention. Referring to Table 1, extensive clinical practice has demonstrated that importing the collected test data from 2529 cases into the early cognitive impairment prediction model provided in this application for prediction showed that the method of this application can achieve a sensitivity of 70% (978 / 1397) and a specificity of 68% (770 / 1132) for MCI screening. Sensitivity (Se): refers to the ability to correctly predict a person who actually has the disease as a patient, i.e., the percentage of patients who are judged as having a positive result. Specificity (Sp): refers to the ability to correctly predict a person who is actually healthy as not having the disease, i.e., the percentage of non-patients who are judged as having a negative result.

[0108] refer to Figure 9 An early cognitive impairment prediction system includes: The data acquisition module 201 is configured to acquire brain bioimpedance data above / below the occipital bone collected by MCI / healthy subjects at multiple preset frequencies. The brain bioimpedance data includes amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope. The optimal single-frequency point determination module 202 is configured to construct a first attention module. It trains the first attention module using brain bioimpedance data collected above / below the occipital bone at multiple preset frequencies and corresponding subject classification labels to obtain a first attention module capable of outputting attention weight sequences for each frequency. Based on the attention weight sequences for each frequency, it determines the optimal single-frequency point for model construction and obtains brain bioimpedance data collected above / below the occipital bone at the optimal single-frequency point. Preferably, the optimal single-frequency point is 50kHz. Preferably, the first attention module is a one-dimensional attention module. During training, it is trained by maximizing the matching degree between the weighted feature sequences of the one-dimensional attention module and the subject classification labels to obtain a first attention module capable of outputting attention weight sequences for each frequency. The risk factor acquisition module 203 is configured to collect multiple basic clinical indicators and physiological parameters related to the subjects, and perform standardization processing to obtain standardized parameters; calculate the standardized mean difference of each standardized parameter between the MCI subject group and the healthy subject group, and sort the standardized parameters according to the standardized mean difference to screen at least one risk factor; preferably, the risk factors include one or more of the following: whether or not the subject has a spouse, whether or not the subject has coronary heart disease, whether or not the subject has hypertension, and whether or not the subject has hyperglycemia. The sample set construction module 204 is configured to construct a training sample set based on brain bioimpedance data collected above / below the occipital bone at the optimal single frequency point and the screened risk factors. The sample set partitioning module 205 is configured to partition the training sample set using a preset mode to obtain six initial sample sets, including: The upper half of the input sample set includes: amplitude, phase, amplitude difference, and phase difference in the occipital region; The lower half of the input sample set includes: suboccipital amplitude, phase, amplitude difference, and phase difference; The dual-semi-input sample set includes: occipital bone amplitude, occipital bone phase, occipital bone amplitude difference, and occipital bone phase difference; The complete input sample set includes: occipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The complete input sample set includes: suboccipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The dual complete input sample set includes: occipital vertical amplitude, occipital vertical phase, occipital vertical amplitude difference, occipital vertical phase difference, amplitude slope, and phase slope. The basic screening model training module 206 is configured to train and cross-validate multiple basic screening models using a multi-layer feedforward neural network for each initial sample set. The early cognitive impairment prediction model construction module 207 is configured to construct an early cognitive impairment prediction model based on the multiple basic screening models and a voting mechanism. Preferably, the early cognitive impairment prediction model construction module is specifically configured to map the output results of each basic screening model to a corresponding risk level. If the output risk levels of all basic screening models are consistent, then that risk level is output. If any basic screening model outputs a high risk and no other basic screening model outputs a low risk, then a high risk is output. If any basic screening model outputs a low risk and no other basic screening model outputs a high risk, then a low risk is output. If there are basic screening models that output both high and low risk, then a medium risk is output. The prediction module 208 is configured to make predictions for the test subjects based on the early cognitive impairment prediction model to obtain prediction results.

[0109] In some embodiments, the risk factor acquisition module is further configured to divide the collected basic clinical indicators and physiological parameters into binary indicators and continuous indicators; wherein the binary indicators are encoded in binary using 0 or 1, and the continuous indicators are converted to the 0-1 range using a normalization method.

[0110] In some embodiments, the early cognitive impairment prediction model building module is further configured to: if the output of a certain basic screening model differs from the output of other models by more than a preset abnormal threshold, then temporarily retrain the basic screening model or temporarily exclude it from the current vote.

[0111] Furthermore, the basic screening model training module is configured to construct male-specific initial sample sets and female-specific initial sample sets for subjects of different genders, and train gender-related basic screening models respectively.

[0112] In some embodiments, the prediction module is further configured to receive personal information of the subject to be tested, the personal information including at least age information; compare the age information with a preset age threshold; if the age information exceeds the preset age threshold, initiate an enhanced screening process, the enhanced screening process including forcibly collecting information on whether the subject to be tested has a spouse, and inputting the information on whether the subject has a spouse as a risk factor into an early cognitive impairment prediction model for prediction.

[0113] In some embodiments, the optimal single-frequency point determination module is specifically configured as follows: A feature vector is constructed from the raw brain bioimpedance data collected above and / or below the occipital bone at each of multiple preset frequency points. ,in Let be the feature dimension at this frequency point, and arrange the feature vectors of all frequency points in frequency order to form a two-dimensional feature matrix. , as input to the attention module; For the input two-dimensional feature matrix Through three weight matrices respectively , and Perform a linear transformation to generate the query matrix. Key matrix Sum matrix : ;in, For the dimensions of query and key, For the dimension of the value, ; Then through the query matrix AND key matrix The dot product is used to calculate the similarity score between each pair of frequency points, and after scaling and softmax normalization, the attention weight matrix is ​​obtained. : ; And the overall attention weight sequence for each frequency point The attention weight matrix is ​​averaged in the value dimension: Among them, the first Line number Column elements Indicates the first The frequency point is paired with the first Attention weights for each frequency point; Then, the attention weight matrix Acting on the value matrix The weighted feature sequence is obtained. : ; Weighted feature sequence Flattened into a one-dimensional vector, it is sequentially input into a fully connected layer and a softmax output layer to obtain the predicted probability that the subject belongs to the MCI category. ; The cross-entropy loss function is used to measure the difference between the predicted probability and the true label. Differences between them: ; With the goal of minimizing cross-entropy loss, the learnable parameters in the attention module are backpropagated and iteratively updated.

[0114] In another aspect, this application also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the following... Figure 1 The method shown.

[0115] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a computer terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0117] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. An early cognitive impairment prediction system, characterized in that, Includes the following steps: The data acquisition module is configured to acquire brain bioimpedance data above / below the occipital bone from a cranial electromagnetic field detection device for early cognitive impairment and healthy subjects at multiple preset frequencies. The brain bioimpedance data includes amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope. The optimal single-frequency point determination module is configured to construct the first attention module. It uses brain bioimpedance data above / below the occipital bone collected by the cranial electromagnetic field detection device at multiple preset frequencies and the corresponding subject classification labels to train the first attention module to obtain a first attention module that can output attention weight sequences at each frequency. The optimal single-frequency point for model construction is determined based on the attention weight sequences at each frequency, and the brain bioimpedance data above / below the occipital bone collected at the optimal single-frequency point is obtained based on the optimal single-frequency point. The risk factor acquisition module is configured to collect multiple basic clinical indicators and physiological parameters related to the subjects, and perform standardization processing to obtain standardized parameters; calculate the standardized mean difference of each standardized parameter between the early cognitive impairment subject group and the healthy subject group, and sort the standardized parameters according to the standardized mean difference to screen at least one risk factor; The sample set construction module is configured to construct a training sample set based on the brain bioimpedance data above / below the occipital bone collected by the cranial electromagnetic field detection device at the optimal single frequency point and the screened risk factors. The sample set partitioning module is configured to partition the training sample set using a preset mode to obtain six initial sample sets, including: The upper half of the input sample set includes: amplitude, phase, amplitude difference, and phase difference in the occipital region; The lower half of the input sample set includes: suboccipital amplitude, phase, amplitude difference, and phase difference; The dual-semi-input sample set includes: occipital bone amplitude, occipital bone phase, occipital bone amplitude difference, and occipital bone phase difference; The complete input sample set includes: occipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The complete input sample set includes: suboccipital amplitude, phase, amplitude difference, phase difference, amplitude slope, and phase slope; The dual complete input sample set includes: occipital vertical amplitude, occipital vertical phase, occipital vertical amplitude difference, occipital vertical phase difference, amplitude slope, and phase slope. The basic screening model training module is configured to train and cross-validate multiple basic screening models using a multi-layer feedforward neural network for each initial sample set. An early cognitive impairment prediction model construction module is configured to construct an early cognitive impairment prediction model based on the multiple basic screening models and the voting mechanism. The prediction module is configured to acquire brain bioimpedance data above / below the occipital bone of the subject under the optimal single frequency point from the cranial electromagnetic field detection device, and input the data into the early cognitive impairment prediction model to predict the subject and obtain a prediction result, which is the risk level of the subject for early cognitive impairment.

2. The early cognitive impairment prediction system according to claim 1, characterized in that, The first attention module is a one-dimensional attention module; Furthermore, during the training process, the first attention module is trained by maximizing the matching degree between the weighted feature sequence of the one-dimensional attention module and the subject's classification label, so as to obtain the first attention module that can output attention weight sequences of each frequency.

3. The early cognitive impairment prediction system according to claim 1, characterized in that, The early cognitive impairment prediction model construction module is specifically configured as follows: Map the output results of each basic screening model to the corresponding risk level; If all basic screening models output the same risk level, then output that risk level. If any basic screening model outputs a high risk and no other basic screening model outputs a low risk, then output a high risk. If any basic screening model outputs low risk and no other basic screening model outputs high risk, then output low risk. If a basic screening model exists that outputs both high-risk and low-risk values, then the output will be medium-risk.

4. The early cognitive impairment prediction system according to claim 1, characterized in that, The risk factor acquisition module is further configured to divide the collected basic clinical indicators and physiological parameters into binary indicators and continuous indicators; wherein, the binary indicators are encoded in binary using 0 or 1, and the continuous indicators are converted to the 0-1 range using a normalization method.

5. The early cognitive impairment prediction system according to claim 3, characterized in that, The early cognitive impairment prediction model construction module is also configured to: if the output of a certain basic screening model differs from the output of other models by more than a preset abnormal threshold, then temporarily retrain the basic screening model or temporarily exclude it from the current vote.

6. An early cognitive impairment prediction system according to any one of claims 1-5, characterized in that, The risk factors include one or more of the following: whether or not one is married, whether or not one has coronary heart disease, whether or not one has hypertension, and whether or not one has hyperglycemia.

7. An early cognitive impairment prediction system according to any one of claims 1-5, characterized in that, The basic screening model training module is configured to construct initial sample sets specifically for males and females for subjects of different genders, and train gender-related basic screening models accordingly.

8. An early cognitive impairment prediction system according to any one of claims 1-5, characterized in that, The prediction module is further configured to receive personal information of the subject to be tested, including at least age information; The age information is compared with a preset age threshold; If the age information exceeds the preset age threshold, an enhanced screening process is initiated. The enhanced screening process includes forcibly collecting information on whether the subject has a spouse, and inputting this information as a risk factor into an early cognitive impairment prediction model for prediction.

9. An early cognitive impairment prediction system according to any one of claims 1-5, characterized in that, The optimal single frequency point is 50KHz.

10. An early cognitive impairment prediction system according to any one of claims 1-5, characterized in that, The optimal single-frequency point determination module is specifically configured as follows: A feature vector is constructed from the raw brain bioimpedance data collected above and / or below the occipital bone at each of multiple preset frequency points. ,in Let be the feature dimension at this frequency point, and arrange the feature vectors of all frequency points in frequency order to form a two-dimensional feature matrix. , as input to the attention module; For the input two-dimensional feature matrix Through three weight matrices respectively , and Perform a linear transformation to generate the query matrix. Key matrix Sum matrix : ;in, For the dimensions of query and key, For the dimension of the value, ; Then through the query matrix AND key matrix The dot product is used to calculate the similarity score between each pair of frequency points, and after scaling and softmax normalization, the attention weight matrix is ​​obtained. : ； And the overall attention weight sequence for each frequency point The attention weight matrix is ​​averaged in the value dimension: Among them, the first Line number Column elements Indicates the first The frequency point is paired with the first Attention weights for each frequency point; Then, the attention weight matrix Acting on the value matrix The weighted feature sequence is obtained. : ; Weighted feature sequence Flattened into a one-dimensional vector, it is sequentially input into a fully connected layer and a softmax output layer to obtain the predicted probability that the subject belongs to the MCI category. ; The cross-entropy loss function is used to measure the difference between the predicted probability and the true label. Differences between them: ； With the goal of minimizing cross-entropy loss, the learnable parameters in the attention module are backpropagated and iteratively updated.

Images