Mild cognitive impairment auxiliary diagnosis method based on motion data analysis

Abstract

The invention provides a mild cognitive impairment auxiliary diagnosis method based on motion data analysis, and relates to the field of cognitive impairment auxiliary diagnosis, and the method specifically comprises the steps: constructing a motion data set, mapping a motion data time sequence to a nonlinear feature space through a hyperbolic tangent function, and obtaining a motion data time sequence; a mean pooling downsampling method is utilized to generate multi-scale sub-time sequence embedding, trend and seasonal components of a multi-scale time sequence are extracted through singular value decomposition, and a projection operator is introduced to construct an adaptive integral equation to realize nonlinear mixing of the trend and seasonal components. According to the method, the mixed trend component and the mixed seasonal component are normalized, the average value of each scale diagnosis value independently predicted by the full-connection neural network is taken to obtain the final mild cognitive impairment auxiliary diagnosis value, and the method can effectively combine multi-level time sequence information to enhance the accuracy of the diagnosis value.

Description

Technical Field

[0001] This invention relates to the field of auxiliary diagnosis of cognitive impairment, and more particularly to an auxiliary diagnostic method for mild cognitive impairment based on motor data analysis. Background Technology

[0002] With the accelerating aging of the population, the incidence of cognitive impairment-related diseases is rising year by year. Mild cognitive impairment, as an important transitional stage between normal aging and dementia, has become a key window for early intervention and risk control of neurodegenerative diseases. If accurate identification and continuous monitoring can be achieved before symptoms are obvious, it will help to carry out intervention measures in advance, slow down the cognitive decline process, reduce the risk of patients transforming into severe cognitive impairment such as Alzheimer's disease, and improve the level of elderly health management and the efficiency of social medical resource utilization.

[0003] Current screening and assessment of mild cognitive impairment mainly rely on methods such as scale assessments and imaging examinations. Among these, scale assessment results are easily affected by subjective factors, cultural background, and testing environment, resulting in limited repeatability and stability. While imaging examinations can provide relatively rich structural or functional information, the equipment is expensive and the examination process is complex, making them unsuitable for large-scale long-term follow-up applications. Overall, existing technologies still have significant shortcomings in the continuous monitoring, objective quantification, and adaptability to daily scenarios for early mild cognitive impairment, making it difficult to meet the needs of home-based and long-term auxiliary diagnosis.

[0004] By collecting and modeling the motor characteristics generated by individuals in their daily activities, we can reflect the potential trend of cognitive function changes from multiple dimensions such as gait rhythm, movement stability and motor coordination. Motor behavior characteristics usually contain information in multiple dimensions. These data show complex and multi-scale changes in time series. Such multi-scale changes require the model to effectively capture short-term and long-term dependencies, so as to give more accurate auxiliary diagnostic results for mild cognitive dysfunction. Summary of the Invention

[0005] This invention proposes an auxiliary diagnostic method for mild cognitive impairment based on motion data analysis. Addressing the shortcomings of existing methods, such as insufficient characterization of the nonlinear temporal features of motion data, inadequate utilization of multi-scale information, and limited ability to model the coupling of trend and periodic changes, this invention constructs a multi-stage mechanism involving multi-scale decomposition, time series decomposition, and multi-scale mixing. This includes mapping the motion data time series to a nonlinear feature space using a hyperbolic tangent function, generating multi-scale sub-time series embeddings through mean pooling downsampling, extracting trend and seasonal components using singular value decomposition, introducing projection operators to calculate dynamic weights to capture information differences between scales, and nonlinearly mixing the trend and seasonal components using an adaptive integral equation to obtain mixed trend and mixed seasonal components. These mixed trend and mixed seasonal components are then normalized using trend and seasonal component action factors, respectively. Finally, a fully connected neural network is used to independently predict the auxiliary diagnostic value for mild cognitive impairment at each scale, and the average value is calculated to obtain the final diagnostic value, thus improving the accuracy and reliability of generating auxiliary diagnostic values ​​for mild cognitive impairment.

[0006] A diagnostic method for mild cognitive impairment based on motor data analysis, characterized by the following steps: S1. Collect motor data that affects cognitive impairment and construct a motor dataset; S2. For the motion data time series in the motion dataset, the hyperbolic tangent function is used to map the motion data time series to the nonlinear feature space to obtain the time series embedding. Mean pooling downsampling is used to process the nonlinear embedding to generate multi-scale sub-time series embeddings. S3. Construct a time series decomposition module to extract the trend and seasonal components embedded in sub-time series through singular value decomposition; S4. Introduce projection operators to calculate the dynamic weights of information at different scales, construct adaptive integral equations, and perform nonlinear mixing of trend and seasonal components to obtain mixed trend components and mixed seasonal components, and construct a multi-scale mixing module. S5. Calculate the first derivatives of the mixed trend and seasonal components respectively, construct the action factors of the trend and seasonal components, and obtain the normalized trend and seasonal component feature vectors through centering and normalization. S6. Using a fully connected neural network, independently predict the auxiliary diagnostic value of mild cognitive impairment for each scale of motion data, including the normalized trend component feature vector and the normalized seasonal component feature vector. Calculate the average value of the diagnostic values ​​to obtain the final auxiliary diagnostic value of mild cognitive impairment. S7. Construct a diagnostic model for mild cognitive impairment, input the motion dataset, and iterate through steps S2 to S6 until convergence.

[0007] Preferably, the construction process of the motion dataset includes two stages: data acquisition and preprocessing. In the data acquisition stage, data such as walking triaxial acceleration, walking triaxial angular velocity, cadence, stride length variation, left and right gait symmetry, average daily exercise duration, activity proportion in different exercise intensity ranges, and continuous static duration generated during daily activities are collected. In the data preprocessing stage, various types of motion data are aligned to a time scale, and the continuously collected motion data are summarized and mapped to the corresponding natural day. Specifically, the statistical characteristics of walking triaxial acceleration and walking triaxial angular velocity are calculated within the daily time range as the motion characteristics of that day. Cadence, stride length variation, and left and right gait symmetry are summarized and processed within the daily scale. Average daily exercise duration, activity proportion in different exercise intensity ranges, and continuous static duration are directly used as the behavioral characteristics of the corresponding day. The motion data time series is constructed using a continuous natural day sequence. All motion data time series together constitute a motion dataset for the auxiliary diagnosis of mild cognitive impairment.

[0008] Preferably, in step S2, the motion data time series in the motion dataset is mapped to a nonlinear feature space using a hyperbolic tangent function to obtain a time series embedding. Mean pooling downsampling is then used to process the nonlinear embedding to generate multi-scale sub-time series embeddings. Linear combination expansion is performed on time series of all types of motion data, and the expansion result is nonlinearly transformed by the hyperbolic tangent function to obtain the time series embedding; The time series embedding is decomposed into... using the mean pooling downsampling method. Sub-time series embeddings at various scales.

[0009] Furthermore, addressing the issues of significant nonlinear evolution in time series motion data during the auxiliary diagnosis of mild cognitive impairment based on motion data analysis, and the difficulty of capturing both subtle local changes and overall evolutionary patterns with a single time scale representation, this study linearly combines and expands all types of motion data time series. This explicitly introduces the linear synergistic relationships and complementary information between different motion data into a unified representation. Then, a hyperbolic tangent function is used to nonlinearly transform the expanded results, making the expanded time series response exhibit stronger separability and a stable numerical range within the nonlinear feature space, thus forming a time series embedding that can be used for subsequent modeling. Furthermore, a mean pooling downsampling method is employed to decompose the time series embedding into... The system generates multi-scale sub-time series embeddings, enabling the short-term fluctuation patterns and long-term trend structures of the same motion behavior at different time resolutions to be preserved and aligned in parallel.

[0010] Preferably, in step S3, a time series decomposition module is constructed to extract the trend and seasonal components of the sub-time series embeddings through singular value decomposition; Singular value decomposition is used to decompose all motion data at all scales into three matrices: a left singular vector matrix, a singular value matrix, and a right singular vector matrix. Before choosing The long-term trend of the sub-time series embedding is extracted by taking the maximum singular value and its corresponding singular vector, and the trend component is obtained. From the Type of motion data The seasonal component is obtained by removing the trend component from the sub-time series embeddings at each scale.

[0011] Furthermore, addressing the challenge of directly separating and modeling the long-term trends and periodic fluctuations of multi-scale sub-time series embeddings in the auxiliary diagnosis of mild cognitive impairment based on motion data analysis, a time series decomposition module is constructed. This module utilizes singular value decomposition to decompose the sub-time series embeddings of all motion data at all scales into three matrices: a left singular vector matrix, a singular value matrix, and a right singular vector matrix. This allows for characterizing the main structural features and energy distribution of the sub-time series embeddings within an orthogonal subspace. Further, by selecting the preceding... The maximal singular value and its corresponding singular vector are used to reconstruct the low-rank structure that dominates the energy in the sub-time series embedding, extract the trend component reflecting the long-term evolution of the motion data, and from the maximal singular value and its corresponding singular vector, the low-rank structure that dominates the energy in the sub-time series embedding is reconstructed, and the trend component reflecting the long-term evolution of the motion data is extracted. Type of motion data By removing the trend component from the sub-time series embeddings at each scale, the corresponding seasonal component is obtained, allowing periodic fluctuations and short-term repetitive patterns to be expressed independently, thereby achieving effective decoupling of long-term trend information and seasonal variation characteristics in the sub-time series embeddings.

[0012] Preferably, in step S4, a projection operator is introduced to calculate the dynamic weights of information at different scales. An adaptive integral equation is constructed by combining the dynamic weights, and the trend and seasonal components are nonlinearly mixed to obtain the mixed trend component and the mixed seasonal component, thus constructing a multi-scale mixing module. Regarding the first By analyzing the trend and seasonal components of various motion data at different scales, a projection operator is constructed that projects from low-scale to high-scale. With the first Using the trend component at the i-th scale as a baseline, adjacent low-scale trend components obtained by the projection operator are introduced and weighted superimposed to obtain the i-th trend component. The first scale and the first The mixed trend component is a weighted mixture of trend components at each scale; With the first Using the seasonal components of the first scale as a baseline, adjacent low-scale seasonal components obtained by the projection operator are introduced and weighted superimposed to construct the first... The first scale and the first The mixed seasonal components are weighted mixtures of seasonal components at each scale.

[0013] Furthermore, addressing the issue that independent modeling of trend and seasonal components at different scales in the auxiliary diagnosis of mild cognitive impairment based on motion data analysis can lead to cross-scale information fragmentation and difficulty in adaptively expressing scale correlations, a projection operator from low-scale to high-scale is constructed for the trend and seasonal components obtained from the c-th type of motion data at different scales to characterize the mapping relationship between features at different time scales; furthermore, using the c-th type of motion data... Using the trend component at the i-th scale as a benchmark, adjacent low-scale trend components obtained by the projection operator are introduced and weighted and superimposed to obtain the i-th trend component. The first scale and the first The mixed trend component is a weighted mixture of trend components at each scale, and simultaneously... Using the seasonal components at the i-th scale as a baseline, adjacent low-scale seasonal components obtained by the projection operator are introduced and weighted superimposed to construct the i-th... The first scale and the first The mixed seasonal components, obtained by weighted mixing of seasonal components at multiple scales, achieve nonlinear fusion of trend information and seasonal information in a multi-scale space. This enables the long-term evolutionary patterns and periodic fluctuation characteristics at different time scales to be synergistically expressed within a unified framework, providing a continuous and consistent multi-scale feature foundation for the stable characterization of dynamic features of subsequent motion data and the auxiliary diagnosis of mild cognitive impairment.

[0014] Preferably, in step S5, the first derivatives of the mixed trend and seasonal components are calculated respectively, the action factors of the trend and seasonal components are constructed, and the normalized trend and seasonal component feature vectors are obtained through centering and normalization. Calculate the first Type of motion data The first of the mixed trend components at each scale The first derivative of the eigenvectors at the time step; For the first derivative of each time step, the square and fourth terms of the mixed trend components of the corresponding time step are combined and weighted and accumulated in the time dimension at preset time intervals to form the trend component action factor. Calculate the first in the time dimension Type of motion data The overall average value of the mixed trend components at each scale is taken and centered. The centering result is then normalized by combining the trend component action factor and the first derivative at the corresponding scale to obtain the normalized trend component feature vector at each time step. With the Type of motion data The first scale mixed trend component The calculation process for the normalized trend component feature vector of the time step feature vector is the same, resulting in the normalized seasonal component feature vector.

[0015] Furthermore, addressing the issue of differences in the temporal evolution intensity, rate of change, and inter-scale amplitude distribution between the mixed trend component and the mixed seasonal component in the auxiliary diagnosis of mild cognitive impairment based on motion data analysis, the calculation of the first... Type of motion data The first of the mixed trend components at each scale The first derivative of the eigenvector at each time step is used to characterize the instantaneous change intensity of the trend component between adjacent time steps. Then, the first derivative of each time step is combined with the square and fourth-order terms of the mixed trend component at the corresponding time step, and weighted and accumulated at preset time intervals in the time dimension to form the trend component action factor. This allows the rate of change and amplitude information of the trend component to be jointly modulated in the time domain. Subsequently, the first derivative is calculated in the time dimension. Type of motion data The overall average of the mixed trend components at each scale is taken and centered. The centered result is then normalized by combining the trend component action factor and first derivative at the corresponding scale, yielding the normalized trend component eigenvectors for each time step. This suppresses the influence of individual differences and amplitude drift on the feature distribution. The method is similar to that used in the first time step. Type of motion data The first scale mixed trend component The normalized trend component feature vector of the time step feature vector is calculated using the same process as the normalized seasonal component feature vector, so that the trend and seasonal components are expressed under a unified numerical scale and dynamic sensitivity, providing a consistent and comparable temporal feature representation for subsequent auxiliary diagnostic models for mild cognitive impairment.

[0016] S6. Using a fully connected neural network, independently predict the auxiliary diagnostic value of mild cognitive impairment for each scale of motion data, including the normalized trend component feature vector and the normalized seasonal component feature vector. Calculate the average value of the diagnostic values ​​to obtain the final auxiliary diagnostic value of mild cognitive impairment. The auxiliary diagnostic value for mild cognitive impairment is calculated independently for each scale of each type of motion data using a fully connected neural network. The average of all diagnostic values ​​is then calculated to obtain the final auxiliary diagnostic value for mild cognitive impairment.

[0017] Furthermore, addressing the issues of varying contributions of different types of motor data, time scales, and trend and seasonal characteristics to the diagnostic results in the auxiliary diagnosis of mild cognitive impairment based on motor data analysis, and the potential bias in decisions based on a single feature or scale, a fully connected neural network is used to independently predict the normalized trend component feature vector and normalized seasonal component feature vector of each type of motor data at each scale, obtaining the corresponding auxiliary diagnostic value for mild cognitive impairment. This allows the influence of various types of motor data and their multi-scale features on the diagnostic results to be independently modeled and fully expressed. Further, by averaging the auxiliary diagnostic values ​​for mild cognitive impairment obtained from all motor data at all scales, the final auxiliary diagnostic value for mild cognitive impairment is obtained. This achieves comprehensive fusion of multi-motor data and multi-scale information at the decision-making level, effectively reducing the impact of abnormal fluctuations in a single feature on the overall diagnostic results and improving the stability, robustness, and consistency of the auxiliary diagnostic results.

[0018] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs an auxiliary diagnostic method for mild cognitive impairment based on motion data analysis. It proposes a multi-stage mechanism that integrates multi-scale decomposition, time series decomposition, and multi-scale mixing. Addressing the shortcomings of existing methods, such as insufficient characterization of temporal nonlinear features of motion data, inadequate utilization of multi-scale information, and limited ability to model the coupling of trends and periodic changes, this invention uses a hyperbolic tangent function to map the time series of motion data to a nonlinear feature space, capturing complex nonlinear relationships. A mean pooling downsampling method is used to generate multi-scale sub-time series embeddings, effectively extracting features at different levels from the data. Then, singular value decomposition is used to extract trend and seasonal components, and a projection operator is introduced to calculate dynamic weights. An adaptive integral equation is combined to achieve nonlinear mixing of trend and seasonal components, further enhancing the dynamic understanding of motion data. Finally, normalization is performed based on the action factors of trend and seasonal components, and a fully connected neural network is used to independently predict the diagnostic value at each scale. The average value of the diagnostic values ​​is calculated, ultimately yielding a more accurate auxiliary diagnostic value for mild cognitive impairment. 2. This invention first extends the time series of motion data in a motion dataset by linearly combining all types of motion data time series to uniformly characterize the correlation and synergistic changes among different motion data. Then, a hyperbolic tangent function is used to perform a nonlinear transformation on the extended result, mapping the original linear changes to a nonlinear feature space to obtain the time series embedding, thereby enhancing the ability to express the nonlinear evolution characteristics of complex motion behaviors. Based on this, the mean pooling downsampling method is further used to process the time series embedding, decomposing the continuous time series embedding into multiple representations with different time resolutions, generating… The sub-time series embedding at various scales enables the model to capture subtle local changes in motion behavior at short time scales and characterize the overall evolution trend of motion patterns at long time scales. 3. This invention utilizes singular value decomposition (SVD) to perform structured analysis of the sub-time series embeddings of all motion data at various scales, decomposing them into three matrices: a left singular vector matrix, a singular value matrix, and a right singular vector matrix. This allows for the characterization of the dominant change direction and energy distribution characteristics of the time series embeddings in algebraic space. Furthermore, by selecting the preceding... The 1st maximal singular value and its corresponding singular vector are used to reconstruct the low-rank structure that occupies the majority of energy in the sub-time series embedding, in order to extract trend components that can reflect the long-term evolution of motion data; based on this, from the 1st... Type of motion data By removing trend components from sub-time series embeddings at various scales, corresponding seasonal components are obtained, enabling periodic fluctuations and short-term recurring patterns to be represented independently. Attached Figure Description

[0019] Figure 1 This is a flowchart of the auxiliary diagnostic method for mild cognitive impairment based on motion data analysis provided by the present invention.

[0020] Figure 2 This is a diagram of the multi-scale time series embedding generation structure provided by the present invention.

[0021] Figure 3 This is a diagram of a multi-scale hybrid module provided by the present invention.

[0022] Figure 4 This is a structural diagram of constructing a time trend sequence of athletic ability provided by the present invention.

[0023] Figure 5 This is a diagnostic effect diagram of the fully connected neural network provided by the present invention. Detailed Implementation

[0024] This invention proposes an auxiliary diagnostic method for mild cognitive impairment based on motion data analysis. Addressing the shortcomings of existing methods, such as insufficient characterization of the nonlinear temporal features of motion data, inadequate utilization of multi-scale information, and limited ability to model the coupling of trends and periodic changes, this invention proposes a multi-stage mechanism integrating multi-scale decomposition, time series decomposition, and multi-scale mixing. This includes using a hyperbolic tangent function to achieve nonlinear mapping of the motion data time series, generating multi-scale sub-time series embeddings through mean pooling downsampling, extracting trend and seasonal components from the sub-time series embeddings using singular value decomposition, introducing a projection operator to calculate dynamic weights, and combining an adaptive integral equation to achieve nonlinear mixing of trend and seasonal components, ensuring effective fusion of information across different scales. Subsequently, normalization is performed by constructing trend and seasonal component action factors to remove scale differences and biases in the data. Finally, a fully connected neural network is used to independently predict the auxiliary diagnostic value for mild cognitive impairment at each scale, and the average of the diagnostic values ​​is calculated to obtain the final auxiliary diagnostic value for mild cognitive impairment.

[0025] S1. Collect motor data that affects cognitive impairment and construct a motor dataset.

[0026] Please see Figure 1 As shown in the embodiments of this application, the auxiliary diagnostic method for mild cognitive impairment based on motion data analysis has the following specific steps.

[0027] In this embodiment, the construction of the motion dataset includes two stages: data acquisition and preprocessing. During the data acquisition stage, data such as walking triaxial acceleration, walking triaxial angular velocity, cadence, stride length variation, left-right gait symmetry, average daily exercise duration, activity proportion in different intensity ranges, and continuous static duration are collected during daily activities. Specifically, walking triaxial acceleration and angular velocity are acquired using wearable inertial sensing devices to reflect the subject's gait rhythm, posture stability, and motion coordination characteristics. Cadence, stride length variation, and left-right gait symmetry are obtained from inertial sensing signal analysis to characterize the consistency of motion execution during walking. Average daily exercise duration, activity proportion in different intensity ranges, and continuous static duration are obtained from the data acquired during daily activities. The statistical results of state recognition are used to characterize the overall activity level and daily motor behavior structure. In the data preprocessing stage, all types of motor data are uniformly aligned to a daily time scale, and the continuously collected motor data are summarized and mapped to the corresponding natural day. Among them, the walking triaxial acceleration and walking triaxial angular velocity are calculated as the motor characteristics of the day within the daily time range. The changes in step frequency, step length, and left and right gait symmetry are summarized and processed within the daily scale. The average daily exercise duration, the proportion of activity in different exercise intensity ranges, and the continuous static duration are directly used as the behavioral characteristics of the corresponding day. The motor data time series is constructed with 7 consecutive natural days. All motor data time series together constitute a motor dataset for the auxiliary diagnosis of mild cognitive impairment.

[0028] S2. For the motion data time series in the motion dataset, the hyperbolic tangent function is used to map the motion data time series to the nonlinear feature space to obtain the time series embedding. Mean pooling downsampling is used to process the nonlinear embedding to generate multi-scale sub-time series embeddings.

[0029] Furthermore, in step S2, sub-time series embeddings are generated, the process of which is as follows: Figure 2 As shown, the specific steps are as follows.

[0030] Linear combination expansion is performed on time series of all types of motion data, and the expansion result is nonlinearly transformed by the hyperbolic tangent function to obtain the time series embedding; In this embodiment, the input number is... Time series of motion data ,in For indexing motion data types, For the number of motion data types, The length of the time series. For the first Type of motion data in the first The collected values ​​at each time step will be used to analyze the motion data in time series. A linear combination extension is performed, mapping to a nonlinear feature space through the hyperbolic tangent function to expand the expressive power of the data. The specific mathematical model is as follows: ; in, Embedding of the mapped time series, For learnable parameter matrix, For the embedded dimension; In this embodiment, the number of motion data types The number of motion data types collected, and the index of the motion data type. Time series length To cover the periodic movements of motion data, embedding dimensions The aim is to expand the original single-channel scalar information into a high-dimensional feature vector with stronger discriminative ability without significantly increasing computational complexity. The time series embedding is decomposed into... using the mean pooling downsampling method. Sub-time series embeddings at various scales; In this embodiment, a mean pooling downsampling method is introduced along the time dimension for the time series embedding, by setting... Using mean aggregation windows of different scales, the time series embeddings are processed by piecewise averaging to obtain... Sub-time series embeddings at various scales are used to reduce data dimensionality and extract features at different scales. The specific mathematical model is as follows: ; in, For the first Type of motion data Sub-time series embeddings at various scales, , For the rounding operation, For mean pooling operation, This is the pooling window size; In this embodiment, the scale value This involves dividing the time series embedding into two different time scales, enabling the model to capture subtle local changes in motion behavior at a short time scale and characterize the overall evolution trend of motion patterns at a long time scale.

[0031] S3. Construct a time series decomposition module to extract the trend and seasonal components embedded in the sub-time series through singular value decomposition.

[0032] Furthermore, in step S3, a time series decomposition module is constructed, and the specific steps are as follows.

[0033] Singular value decomposition is used to decompose all motion data at all scales into three matrices: a left singular vector matrix, a singular value matrix, and a right singular vector matrix. In this embodiment, the first Type of motion data in the first Sub-time series embeddings at various scales Treating it as a two-dimensional feature matrix, performing singular value decomposition yields three matrices: a left singular vector matrix, a singular value matrix, and a right singular vector matrix. The specific mathematical model is as follows: ; in, It is a left singular vector matrix. It is a singular value matrix. It is a right singular vector matrix. This is the transpose operation of a matrix; Before choosing The long-term trend of the sub-time series embedding is extracted by taking the maximum singular value and its corresponding singular vector, and the trend component is obtained. In this embodiment, for the first Type of motion data Sub-time series embeddings at various scales After completing the singular value decomposition, select the top-ranking singular values ​​in order of their magnitude. Each corresponding component is extracted from the left singular vector matrix, singular value matrix, and right singular vector matrix, respectively, and the components corresponding to the preceding components are extracted. The vectors corresponding to the singular values ​​and their diagonal elements are used to reconstruct the truncated matrices according to their original product relationships, forming the nth singular value. Type of motion data in the first The trend components of sub-time series at various scales are represented by the following mathematical model: ; in, For the first Type of motion data Trend components at each scale, From Extracting the first A left singular vector From Extracted from the front A diagonal matrix composed of singular values. From Extracting the first One right singular vector; From the Type of motion data The seasonal component is obtained by removing the trend component from the sub-time series embeddings at each scale; In this embodiment, the first Type of motion data Sub-time series embeddings at various scales As a baseline, subtract the first Type of motion data By embedding the trend component of each scale sub-time series, the seasonal component is obtained. The specific mathematical model is as follows: ; in, For the first Type of motion data Seasonal components embedded in sub-scale time series.

[0034] S4. Introduce projection operators to calculate the dynamic weights of information at different scales, construct adaptive integral equations, and perform nonlinear mixing of trend and seasonal components to obtain mixed trend components and mixed seasonal components, thus constructing a multi-scale mixing module.

[0035] Furthermore, in step S4, a multi-scale hybrid module is constructed, the process of which is as follows: Figure 3 As shown, the specific steps are as follows.

[0036] In this embodiment, the first... Type of motion data Based on the trend and seasonal components at each scale, a weighted projection and recombination of the low-scale trend and seasonal components is performed using a weighting method based on exponential normalization, forming a structure that is similar to the first-scale trend and seasonal components in terms of time length and feature dimensions. A projection operator with consistent scale, the specific mathematical model of which is: ; ; in, , The first Trends and seasonal components of various sports data from the first Scale projection to the first The projection operator of the scale, where , , , , For a trainable parameter matrix, , The first Type of motion data Scale trends and seasonal components, This is the transpose operation of a matrix; With the first Using the trend component at the i-th scale as a baseline, adjacent low-scale trend components obtained by the projection operator are introduced and weighted superimposed to obtain the i-th trend component. The first scale and the first The mixed trend component is a weighted mixture of trend components at each scale; In this embodiment, the first Using the trend components at the first scale as a baseline, adjacent low-scale trend components obtained by the projection operator are introduced and continuously weighted and accumulated over time, so that the low-scale trend components gradually decay with time distance before being superimposed to the first scale. From the trend components at each scale, a mixed trend component is obtained. The specific mathematical model is as follows: ; in, For the first Type of motion data The first scale and the first The mixed trend component is a weighted mixture of trend components at each scale. For learnable hyperparameters that control the decay of trend component information; In this embodiment, the hyperparameter controlling the decay of trend component information is... Initialize it to 0.001 to keep the information decay over time gradual; With the first Using the seasonal components of the first scale as a baseline, adjacent low-scale seasonal components obtained by the projection operator are introduced and weighted superimposed to construct the first... The first scale and the first The mixed seasonal components are weighted and mixed at each scale; In this embodiment, the first Using the seasonal components at the lower scale as a benchmark, adjacent lower-scale seasonal components obtained by the projection operator are introduced and continuously weighted and accumulated over time, so that the lower-scale seasonal components gradually decay with time distance before being superimposed to the next scale. The specific mathematical model for the seasonal components at each scale is as follows: ; in, For the first Type of motion data Scale and the first The mixed seasonal components after weighted mixing of seasonal components at different scales, For learnable hyperparameters to control the decay of seasonal component information In this embodiment, the hyperparameter controlling the decay of seasonal component information is... Initialize it to 0.001 to keep the information decay over time gradual.

[0037] S5. Calculate the first derivatives of the mixed trend and seasonal components respectively, construct the action factors of the trend and seasonal components, and obtain the normalized trend and seasonal component feature vectors through centering and normalization.

[0038] Furthermore, in step S5, the normalized trend and seasonal component eigenvectors are calculated, and the specific steps are as follows.

[0039] Calculate the first Type of motion data The first of the mixed trend components at each scale The first derivative of the eigenvectors at the time step; In this embodiment, for the first Type of motion data in the first The mixed trend components obtained at each scale are calculated by differencing the feature vectors of adjacent time steps along the time dimension, and then normalized by incorporating the time interval between adjacent time steps, thus obtaining the first derivative. The specific mathematical model is as follows: ; in, For the first Type of motion data The first of the mixed trend components at each scale The first derivative of the eigenvectors at each time step. For the first Type of motion data The first of the mixed trend components at each scale The feature vector at each time step For the first Type of motion data The first of the mixed trend components at each scale The feature vector at each time step for , The time step interval between feature vectors; In this embodiment, the time step interval This preserves the original characteristics of the features while avoiding excessive amplification of subtle fluctuations. For the first derivative of each time step, the square and fourth terms of the mixed trend components of the corresponding time step are combined and weighted and accumulated in the time dimension at preset time intervals to form the trend component action factor. In this embodiment, the weighted cumulative number of steps is calculated according to a preset time step interval. Type of motion data in the first The effect factor of the trend component is obtained by taking the squares of the first derivatives of all time steps at each scale, the squares of the mixed trend component, and the quartic term. The specific mathematical model is as follows: ; in, For the first Type of motion data in the first The influence factor of the trend component at each scale. , These are the hyperparameters that represent the local and global effects of learnable control data, respectively; In this embodiment, the hyperparameter controls the local influence of data. Initializing it to 0.1 makes the model sensitive to slight local changes while avoiding the dominant effect of local perturbations on the overall trend estimation; this is a hyperparameter that controls the global impact of the data. Initializing it to 0.01 helps suppress nonlinear anomalies with large global amplitudes and prevents the model from overfitting to unrepresentative long-period perturbations. Calculate the first in the time dimension Type of motion data The overall average value of the mixed trend components at each scale is taken and centered. The centering result is then normalized by combining the trend component action factor and the first derivative at the corresponding scale to obtain the normalized trend component feature vector at each time step. In this embodiment, the first step is calculated in the time dimension. Type of motion data The overall average value of the mixed trend components at each scale is calculated, and adjusted by a trend component normalization scaling factor. This adjusted overall average value is then subtracted from the trend component feature vector at the corresponding time step to eliminate the global bias effect, thus completing the centering process. Further, combining the trend component action factor and the first derivative, a normalized denominator is constructed, and the centering result is normalized to obtain the normalized trend component feature vector. The specific mathematical model is as follows: ; in, For the first Type of motion data Scale mixing trend component The normalized trend component eigenvector of the time step eigenvector. The normalization scaling factor is used to normalize the learnable trend components. for The absolute value; In this embodiment, the trend component normalization scaling factor Initialize to 0.5 to maintain a balance between current and historical information while preserving the differences in feature vectors; With the Type of motion data The first scale mixed trend component The calculation process for the normalized trend component feature vector of the time step feature vector is the same, resulting in the normalized seasonal component feature vector. In this embodiment, with the first Type of motion data The first scale mixed trend component The calculation process for the normalized trend component feature vector of the time step feature vector is the same, yielding the normalized seasonal component feature vector. The specific mathematical model is as follows: ; in, For the first Type of motion data The seasonal component at each scale Normalized seasonal component feature vectors of eigenvectors, For seasonal components, normalization scaling factor, For the first Type of motion data in the first The effect factor of seasonal components at each scale, For the first Type of motion data The first of the seasonal components at each scale The feature vector at each time step For the first Type of motion data The first of the seasonal components at each scale The first derivative of the eigenvectors at each time step. for The absolute value; In this embodiment, the seasonal component normalization scaling factor Initialize it to 0.5 to maintain a balance between current and historical information while preserving the differences in feature vectors.

[0040] S6. Using a fully connected neural network, independently predict the auxiliary diagnostic value of mild cognitive impairment for each scale of each type of motion data, including the normalized trend component feature vector and the normalized seasonal component feature vector. Calculate the average value of the diagnostic values ​​to obtain the final auxiliary diagnostic value of mild cognitive impairment.

[0041] Furthermore, in step S6, the auxiliary diagnostic value for mild cognitive impairment is calculated, and the procedure is as follows: Figure 4 As shown, the specific steps are as follows.

[0042] The auxiliary diagnostic value of mild cognitive impairment is calculated independently for each scale of each type of motion data using a fully connected neural network. The average of all diagnostic values ​​is then calculated to obtain the final auxiliary diagnostic value of mild cognitive impairment. In this embodiment, an independent fully connected neural network is constructed to map the normalized trend component feature vector and the normalized seasonal component feature vector at each scale for each type of motion data, obtaining the corresponding primary diagnostic results. Subsequently, all primary diagnostic results are weighted and aggregated, and the diagnostic values ​​of all motion data are averaged along the motion data dimension to finally form an auxiliary diagnostic value for mild cognitive impairment. The specific mathematical model is as follows: ; in, This serves as the final auxiliary diagnostic value for mild cognitive impairment. , The first Type of motion data A fully connected neural network for scale trends and seasonal components. For the first Type of motion data Normalized trend component feature sequences at each scale, , For the first Type of motion data Each scale-normalized seasonal component feature sequence .

[0043] To address the shortcomings of existing methods in characterizing the nonlinear features of motion data, fully utilizing multi-scale information, and limiting the ability to model coupled trend and periodic changes, a multi-stage mechanism is constructed, involving multi-scale decomposition, time series decomposition, and multi-scale mixing. When a motion dataset is input, the mechanism sequentially performs stages such as multi-scale sub-time series embedding, trend and seasonal component analysis, mixing trend and seasonal components, and generating normalized trend and seasonal component feature vectors. During training, a loss function is adopted... The loss function achieves model convergence through iterative training and adaptive updating of model parameters, ultimately resulting in an auxiliary diagnostic model for mild cognitive impairment based on motion data analysis.

[0044] Furthermore, in step S6, the mild cognitive impairment auxiliary diagnostic model based on motion data analysis proposed in this invention is implemented using the Python 3.8 programming language and the PyTorch framework. To improve training efficiency, the method runs in a CUDA 11 environment and uses an NVIDIA A100 GPU for computation. During training, the initial learning rate is set to... Batch size is ,use The optimizer performs gradient updates, and the loss function is selected. Loss function.

[0045] Furthermore, in step S6, the motion dataset is input into the constructed model for training and evaluation. The diagnostic performance of this method is as follows: Figure 5 As shown, the vertical axis represents the auxiliary diagnostic value for mild cognitive impairment, and the horizontal axis represents time. The gray dashed line in the figure represents the actual auxiliary diagnostic value for mild cognitive impairment, while the black solid line represents the auxiliary diagnostic value for mild cognitive impairment obtained by this method. Figure 5 It can be seen that the diagnostic values ​​obtained by this method have a high degree of fit with the actual diagnostic values, indicating that this method can effectively assist in the diagnosis of mild cognitive impairment and has good accuracy and stability.

[0046] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for assisting in the diagnosis of mild cognitive impairment based on motor data analysis, characterized in that, Includes the following steps: S1. Collect motor data that affects cognitive impairment and construct a motor dataset; S2. For the motion data time series in the motion dataset, the hyperbolic tangent function is used to map the motion data time series to the nonlinear feature space to obtain the time series embedding. Mean pooling downsampling is used to process the nonlinear embedding to generate multi-scale sub-time series embeddings. S3. Construct a time series decomposition module to extract the trend and seasonal components embedded in sub-time series through singular value decomposition; S4. Introduce projection operators to calculate the dynamic weights of information at different scales, construct adaptive integral equations, and perform nonlinear mixing of trend and seasonal components to obtain mixed trend components and mixed seasonal components, and construct a multi-scale mixing module. S5. Calculate the first derivatives of the mixed trend and seasonal components respectively, construct the action factors of the trend and seasonal components, and obtain the normalized trend and seasonal component feature vectors through centering and normalization. S6. Using a fully connected neural network, independently predict the auxiliary diagnostic value of mild cognitive impairment for each scale of motion data, including the normalized trend component feature vector and the normalized seasonal component feature vector. Calculate the average value of the diagnostic values ​​to obtain the final auxiliary diagnostic value of mild cognitive impairment. S7. Construct a diagnostic model for mild cognitive impairment, input the motion dataset, and iterate through steps S2 to S6 until convergence.

2. The auxiliary diagnostic method for mild cognitive impairment based on motion data analysis according to claim 1, characterized in that, The construction of the motion dataset includes two stages: data acquisition and preprocessing. In the data acquisition stage, data on walking triaxial acceleration, walking triaxial angular velocity, cadence, stride length variation, left and right gait symmetry, average daily exercise duration, activity proportion in different exercise intensity ranges, and continuous static duration generated during daily activities are collected. In the data preprocessing stage, various types of motion data are aligned to a time scale, and the continuously collected motion data are summarized and mapped to the corresponding natural day. Specifically, the statistical characteristics of walking triaxial acceleration and walking triaxial angular velocity are calculated within the daily time range as the motion characteristics of that day. Cadence, stride length variation, and left and right gait symmetry are summarized within the daily scale. Average daily exercise duration, activity proportion in different exercise intensity ranges, and continuous static duration are directly used as the behavioral characteristics of the corresponding day. The motion data time series is constructed using a continuous natural day sequence. All motion data time series together constitute a motion dataset for the auxiliary diagnosis of mild cognitive impairment.

3. The auxiliary diagnostic method for mild cognitive impairment based on motion data analysis according to claim 2, characterized in that, Linear combination expansion is performed on time series of all types of motion data, and the expansion result is nonlinearly transformed by the hyperbolic tangent function to obtain the time series embedding; The time series embedding is decomposed into... using the mean pooling downsampling method. Sub-time series embeddings at various scales.

4. The auxiliary diagnostic method for mild cognitive impairment based on motion data analysis according to claim 3, characterized in that, Singular value decomposition is used to decompose all motion data at all scales into three matrices: a left singular vector matrix, a singular value matrix, and a right singular vector matrix. Before choosing The long-term trend of the sub-time series embedding is extracted by taking the maximum singular value and its corresponding singular vector, and the trend component is obtained. From the Type of motion data The seasonal component is obtained by removing the trend component from the sub-time series embeddings at each scale.

5. The auxiliary diagnostic method for mild cognitive impairment based on motion data analysis according to claim 4, characterized in that, Regarding the first By analyzing the trend and seasonal components of various motion data at different scales, a projection operator is constructed that projects from low-scale to high-scale. With the first Using the trend component at the i-th scale as a baseline, adjacent low-scale trend components obtained by the projection operator are introduced and weighted and superimposed to obtain the i-th trend component. The first scale and the first The mixed trend component is a weighted mixture of trend components at each scale; With the first Using the seasonal components of the first scale as a baseline, adjacent low-scale seasonal components obtained by the projection operator are introduced and weighted superimposed to construct the first... The first scale and the first The mixed seasonal components are weighted mixtures of seasonal components at each scale.

6. The auxiliary diagnostic method for mild cognitive impairment based on motion data analysis according to claim 5, characterized in that, Calculate the first Type of motion data The first of the mixed trend components at each scale The first derivative of the eigenvectors at the time step; For the first derivative of each time step, the square and fourth terms of the mixed trend components of the corresponding time step are combined and weighted and accumulated at preset time intervals in the time dimension to form the trend component action factor.

7. The auxiliary diagnostic method for mild cognitive impairment based on motion data analysis according to claim 6, characterized in that, Calculate the first in the time dimension Type of motion data The overall average value of the mixed trend components at each scale is taken and centered. The centering result is then normalized by combining the trend component action factor and the first derivative at the corresponding scale to obtain the normalized trend component feature vector at each time step. With the Type of motion data The first scale mixed trend component The calculation process for the normalized trend component feature vector of the time step feature vector is the same, resulting in the normalized seasonal component feature vector.

8. The auxiliary diagnostic method for mild cognitive impairment based on motion data analysis according to claim 7, characterized in that, The auxiliary diagnostic value for mild cognitive impairment is calculated independently for each scale of each type of motion data using a fully connected neural network. The average of all diagnostic values ​​is then calculated to obtain the final auxiliary diagnostic value for mild cognitive impairment.

Images