MCI detection system based on dynamic adaptive fusion of physiological signals and cognitive data

Abstract

The invention discloses an MCI detection system based on dynamic adaptive fusion of physiological signals and cognitive data. Comprising a data acquisition module, a time sequence splitting and aligning module, a multi-modal data preprocessing module, a multi-scale feature extraction module, a dynamic adaptive feature selection module based on harmony search and a dynamic adaptive stacking classification module based on harmony search. According to the dynamic adaptive feature selection optimization algorithm based on harmony search, redundant information is reduced, and a feature set is optimized. And finally, by proposing a dynamic self-adaptive stacking classification optimization algorithm based on harmony search, accurate classification of MCI patients and healthy individuals is realized. According to the method, the conventional wearable device is combined with cognitive evaluation on the tablet computer, the method has the advantages of being easy and convenient to operate, suitable for the family environment and the like, and important technical support is provided for autonomously detecting the MCI.

Description

Technical Field

[0001] The present invention relates to the technical field of mild cognitive impairment detection in medical diagnosis, and in particular to an MCI detection system based on dynamic adaptive fusion of physiological signals and cognitive data. Background Art

[0002] Mild cognitive impairment (MCI) is widely considered a precursor to dementia, representing an intermediate stage between normal cognitive aging and the severe pathological decline seen in dementia. Although various potential treatments for dementia have been proposed, such as inhibiting enzymes associated with amyloid-β production and using antibodies to clear amyloid-β from the brain, there is currently no drug that can completely cure dementia or significantly alter its clinical course. At the same time, studies have shown that early intervention is most effective, which requires accurate and sensitive diagnosis of MCI. Therefore, effective MCI detection is crucial, as this may enable patients to receive timely interventions that can slow cognitive decline and alleviate the burden of dementia.

[0003] At present, the rich individual physiological data collected by wearable devices provide a new detection method for mining the physiological and behavioral characteristics of patients with cognitive impairment, thereby realizing timely, economical and efficient detection of mild cognitive impairment (MCI). Studies have shown that physiological indicators such as heart rate variability, skin electrodermal activity, gait variability, skin temperature, respiratory rate, electroencephalogram, eye tracking and myoelectric activity can effectively reflect changes in cognitive function and provide an objective basis for early diagnosis of cognitive impairment. However, due to the diversity and individual differences of these physiological indicators, the selection of appropriate monitoring parameters becomes particularly complicated, which increases the difficulty for doctors and patients to make decisions using physiological monitoring data. In addition, the existing cognitive function monitoring practice mainly focuses on the analysis of physiological parameters, but lacks a comprehensive analysis of these physiological data with traditional or computerized cognitive test results, which limits the comprehensive evaluation of cognitive function and does not give full play to the complementary potential of physiological data and cognitive test results. Therefore, how to efficiently integrate and evaluate diverse physiological measurement data and adaptively integrate indicators with significant monitoring value with cognitive test measurement parameters to achieve comprehensive cognitive function evaluation has become an urgent problem to be solved. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes an MCI detection system based on the dynamic adaptive fusion of physiological signals and cognitive data, thereby improving the accuracy and practicality of MCI detection. In this system, firstly, the time series alignment of skin electrical activity, photoplethysmography and digital cognitive parameters is performed by cubic spline interpolation and timestamp sorting matching technology. Secondly, a multi-stage automatic artifact removal algorithm is designed, which combines artifact correction, signal decomposition and overlapping sliding time window technology to perform data purification. Based on this, multi-scale features are extracted by time domain, frequency domain and nonlinear analysis methods. Then, a dynamic adaptive feature selection algorithm based on harmony search is proposed, which selects the feature subset that is most helpful for classification by adaptively updating the harmony memory library, and introduces adaptive hyperparameter adjustment to iteratively improve classification performance. The present invention provides a more accurate and practical detection method for the early diagnosis of MCI by innovatively combining physiological data with cognitive tests.

[0005] The technical solution of the present invention is as follows:

[0006] An MCI detection system based on dynamic adaptive fusion of physiological signals and cognitive data, characterized by comprising a data acquisition module, a time series splitting and alignment module, a multimodal data preprocessing module, a multi-scale feature extraction module, a dynamic adaptive feature selection module based on harmony search, and a dynamic adaptive stacking classification module based on harmony search;

[0007] Data collection module: The subjects first wore Empatica 4 on their non-dominant wrist and sat quietly with their eyes closed for five minutes to collect physiological baseline signals. Then, the subjects performed cognitive tasks on iPad 2019 and used Apple Pencil to complete drawing-related operations, and their cognitive performance and physiological signals were recorded synchronously throughout the process;

[0008] Time series splitting and alignment module: Based on the heterogeneity of the collected data (such as EDA, IBI and cognitive parameters), the present invention proposes EDA time series processing and IBI time series processing strategies to maintain spatiotemporal continuity and effectively deal with potential signal discontinuities, inconsistencies, and misalignments between existing data;

[0009] Multimodal data preprocessing module: used to process multimodal data that remain consistent in time series, including EDA signal preprocessing, IBI signal preprocessing and digital cognitive parameter preprocessing:

[0010] The data preprocessing module comprises:

[0011] EDA signal preprocessing unit: used for EDA signal artifact correction, signal decomposition and overlapping sliding time window processing;

[0012] IBI signal preprocessing unit: used for IBI signal artifact removal, cubic spline interpolation and overlapping sliding time window processing;

[0013] Digital cognitive parameter preprocessing unit: used to fill in missing values, remove outliers, and check data consistency and accuracy for the mean of digital cognitive parameters;

[0014] Multi-scale feature extraction module: for pre-processed multimodal data, the present invention constructs feature vectors for EDA signals and HRV signals based on the time domain, constructs feature vectors for EDA signals and HRV signals based on the frequency domain, and constructs feature vectors based on digital cognitive parameters;

[0015] Dynamic adaptive feature selection module based on harmony search: For the extracted high-dimensional feature set, the present invention adopts an improved harmony search algorithm, analogizes a tone of each instrument to a feature, dynamically adjusts the hyperparameters in each iteration process, so that the randomly generated feature solution vector changes, and finally obtains a feature subset with lower dimension that is suitable for most base classifiers;

[0016] Dynamic adaptive stacking classification module based on harmony search: The present invention improves the harmony search algorithm, makes full use of the complementarity and correlation between multimodal data, compares each basic learner to a tone, and dynamically adjusts the hyperparameters according to the accuracy and number of learners, ultimately achieving accurate classification of MCI patients.

[0017] The system and the EDA time series processing strategy described are as follows: Since the EDA.csv file downloaded from the cloud only contains the session start time and sampling rate information, and lacks the timestamp corresponding to each signal point, the present invention designs a timestamp completion strategy based on the sampling rate. According to the session start time and the sampling rate of 4Hz, a timestamp is automatically generated for every four data points. Then, the entire EDA time series is divided into a baseline data segment and a test data segment, and the baseline segment is defined as the first 5 minutes after the start of the session. Finally, the timestamp of the cognitive task is aligned with the EDA sequence timestamp, and the EDA signal segment during the test is extracted.

[0018] The system, the IBI time series processing strategy: The photoelectric sensor of the Empatica 4 wristband will automatically discard the heartbeat signal that cannot be accurately identified when collecting IBI data, resulting in discontinuity of the data in the IBI.csv file. Therefore, the present invention adopts a missing data completion method for the IBI time series. First, the physiological reasonable range of IBI is limited to within 2 seconds, and any data beyond this range will be marked as missing. Subsequently, these gaps are estimated using cubic spline interpolation to ensure the integrity of the time series. The time continuity of the IBI data is optimized by the curve fitting method, and finally the IBI data is synchronized with the EDA data, timestamps are added, baseline and test data segments are divided, and aligned with the timestamps of cognitive tasks.

[0019] The system, the EDA signal artifact correction: first, use a first-order Butterworth low-pass filter with a cutoff frequency of 0.6 Hz to filter the EDA signal, retain the low-frequency components in the EDA signal, and remove high-frequency noise; the transfer function formula of the filter is:

[0020]

[0021] Among them, s is a complex variable, ω c is the cutoff frequency.

[0022] Secondly, the EDAexplorer tool is used to detect artifacts and mark the artifacts of the EDA data after low-pass filtering; then, the artifact signal is interpolated using the cubic spline interpolation method, and the missing data points are approximated by piecewise cubic polynomials to ensure that the function value, first-order derivative, and second-order derivative at each data point are continuous. Suppose there is a set of EDA data (x1, y1), (x2, y2), ..., (x i ,y i ), some of the data points have gaps that need to be interpolated. In each gap interval [x i ,x i+1 ], using a cubic polynomial S i (x) is interpolated:

[0023] S i (x) = a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3

[0024] Among them, a i ,b i ,ci ,d i are the unknown coefficients, which are determined by satisfying the interpolation conditions, including at each data point x i At this point, the function value of the interpolation polynomial is equal to the ordinate value of the data point, that is, S i (x i )=y i ; At each data point x i At , the first-order derivative of the interpolation polynomial is continuous, that is, S i '(x i )=S i ' +1 (x i ); at each data point x i At , the second-order derivative of the interpolation polynomial is continuous, that is, S i ”(x i )=S i ” +1 (x i ).

[0025] The system, the EDA signal decomposition: cvxEDA regards the EDA signal as a superposition of Tonic components, Phasic components and additive white Gaussian noise by solving a convex optimization problem. The optimization problem is reformulated as a standard quadratic optimization (i.e., the objective function is a quadratic polynomial of the variables, and the constraints are affine functions), so as to achieve efficient solution while considering the constraints. By minimizing the loss function, cvxEDA makes the decomposed Phasic and Tonic components fit the original signal as much as possible and minimizes the impact of noise. This optimization problem is defined as:

[0026]

[0027] subject to Aq≥0

[0028] Among them, M and A are tridiagonal matrices; q is an auxiliary variable; B is a matrix whose columns are cubic B-spline basis functions; is the spline coefficient vector; C is an N×2 matrix; d is a 2×1 vector containing the offset and slope coefficients of the linear trend; y is a given N sample length EDA signal; parameters α and γ control the penalty intensity for the Tonic and Phasic components respectively. The larger α is, the sparser the estimation result is, and most of the spurious spikes caused by noise will be suppressed, but the signal distortion (that is, the attenuation of true activation) will also be greater. Conversely, the smaller α is, the smaller the distortion is, but the greater the noise is. Regarding γ, the larger the value, the stronger the penalty for , that is, the smoother the Tonic curve.

[0029] After finding the optimal [q, l, d], the Phasic component can be obtained by p = Mq, while the Tonic component can be easily obtained by t = B + Cd. If there is physiological incoherence data, the Tonic and Phasic components can be further filtered using a low-pass Butterworth filter to remove negative SCR and SCL values.

[0030] The system, the EDA signal overlap sliding time window processing: In view of the minimum time of cognitive test, the present invention sets the window size to 60 seconds, the step size of the sliding window is 1 second, and a time series is generated according to the first timestamp and the last timestamp of the EDA data. This sequence will be used as the starting point of the time window. Assume that the first timestamp is t1 and the last timestamp is t n , the window size is w, and the step size of the sliding window is s, then the value of the time series is [t1, t1+s, t1+2×s, t1+3×s, ..., t n -w].

[0031] The system, the IBI signal artifact removal: four artifact detection rules are used to identify IBI artifacts obtained from the Empatica 4 wristband, and the formulas are:

[0032] IBI i ≥250ms IBI i ≤2000ms

[0033] |IBI i -IBI i-1 |≤0.2×IBI i-1

[0034]

[0035]

[0036] The system, the IBI signal cubic spline interpolation and overlapping sliding time window processing: applying a cubic spline interpolation method similar to the EDA signal to the detected IBI artifact signal to fill in missing values, and at the same time, dividing the processed IBI signal into overlapping time windows of 60 seconds in size and 1 second in step size, thereby generating continuous data for subsequent feature extraction;

[0037] The system, the feature vectors of EDA signal and HRV signal are constructed based on time domain: by calculating the SCL (Tonic component) and SCR (Phasic component) of each window, the mean, standard deviation, median, root mean square, minimum, maximum, energy, skewness, kurtosis, sample entropy and other EDA time domain features are extracted; at the same time, by calculating the IBI sequence, the mean, median, standard deviation, root mean square deviation of the IBI interval, the number of consecutive IBI intervals with a difference of more than 50 milliseconds and the corresponding percentage value, as well as the mean, standard deviation, minimum and maximum of the heart rate and other HRV time domain features are extracted;

[0038] The system described herein constructs a feature vector for an EDA signal based on the frequency domain: the Welch method is used to estimate the power spectral density (PSD), and the Tonic component and the Phasic component are subdivided into five frequency bands within the recommended frequency range (i.e., 0.05-0.50 Hz). The Welch method is a frequency domain analysis technique that divides a time series into multiple overlapping windows, and then applies a Fourier transform to each window to obtain a PSD estimate for each window. The PSD of each window obtained in this way represents the frequency domain characteristics of the signal in different time periods. Subsequently, the PSD values ​​of all windows are averaged to obtain an overall PSD estimate. On this basis, for each predefined frequency band, the statistical characteristics of its spectral power, such as the mean, standard deviation, median, kurtosis, skewness, maximum value, and minimum value, are calculated.

[0039] In addition, the HRV signal is decomposed in the frequency domain according to different frequency band ranges, and is divided into three frequency bands: high frequency (HF, 0.15Hz to 0.4Hz), low frequency (LF, 0.04Hz to 0.15Hz) and very low frequency (VLF, 0.003Hz to 0.04Hz). The power spectral density (PSD) is estimated by the Lomb-Scargle method, and characteristic parameters such as very low frequency power, low frequency power, high frequency power, total power and LF / HF ratio are extracted from it. The formula of the Lomb-Scargle periodogram method is as follows:

[0040]

[0041] Among them, ω j =2πf j , f j is each frequency to be tested, N is the number of data points, t j is the timestamp of each data point, y(t j ) is at time t j The measured value of is the mean value of y(t), σ is the variance of y(t), and τ is the time offset.

[0042] The system described herein constructs a feature vector based on digital cognitive parameters: for cognitive tasks, it includes the time and score for completing each task, as well as handwriting, frequency, curvature and sequence features related to the drawing task.

[0043] The system, for dynamic adaptive feature selection based on harmony search, introduces the Hamming distance to measure the difference between the currently generated harmony vector and the optimal vector in the harmony memory library. The Hamming distance is used to evaluate the similarity between the two vectors to adjust the search probability. When the Hamming distance is high, the search probability is moderately reduced to enhance the exploratory nature; conversely, the search probability is moderately increased to enhance the tendency to utilize known information. At the same time, the present invention also optimizes the fitness function value of the harmony vector by using ensemble learning, takes the average performance of all base classifiers as the target, and combines the feature selection rate to improve and balance the classification effect, in order to obtain a feature subset with a lower dimension that is suitable for most base classifiers.

[0044] The system, for dynamic adaptive stacked classification based on harmony search, defines the fitness function value of the harmony vector as a combined function of the classification error rate and the number of base learners, and continuously iterates the fitness function value of the harmony vector according to the selection probability of the harmony memory library to minimize it, thereby obtaining the final MCI detection result.

[0045] The beneficial effects of the present invention are:

[0046] The method of the present invention constructs a dynamic adaptive integrated learning system based on harmony search optimization, aiming to integrate data from wearable wristbands and tablet devices to establish a comprehensive and reliable MCI detection method. In response to the problem of feature redundancy, the present invention introduces an improved harmony search feature selection algorithm, which effectively avoids the trap of local optimal solutions through an adaptive hyperparameter adjustment strategy, and establishes a set of comprehensive evaluation indicators to extract a more comprehensive feature subset on each basic learner. Then, the system uses nine heterogeneous learners as basic learners, and uses the improved harmony search algorithm to combine and optimize these learners, thereby constructing a dynamic adaptive stacked classification model. The results of the present invention have the remarkable characteristics of high cost-effectiveness, excellent diagnostic accuracy and good transferability. Its simple and easy-to-use characteristics make it show a wide range of application potential in clinical auxiliary diagnosis in the field of neurocognitive disorders, thereby providing more accurate and personalized diagnosis and treatment solutions for the medical field. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of a dynamic adaptive MCI detection system based on multi-source information fusion of the present invention;

[0048] Figure 2It is a schematic diagram of the splitting and time alignment process based on EDA data of the present invention;

[0049] Figure 3 Schematic diagram of the splitting and time alignment process based on IBI data of the present invention;

[0050] Figure 4 This is a schematic diagram of the EDA signal preprocessing process of the present invention;

[0051] Figure 5 This is a schematic diagram of the IBI signal preprocessing process of the present invention;

[0052] Figure 6 A schematic diagram showing a comparison of evaluation indicators between the model proposed in the present invention and a machine learning model using a dynamic adaptive feature selection algorithm based on harmony search (AHSI). DETAILED DESCRIPTION

[0053] The present invention is described in detail below in conjunction with specific embodiments.

[0054] An MCI detection system based on dynamic adaptive fusion of physiological signals and cognitive data, characterized by comprising a data acquisition module, a time series splitting and alignment module, a multimodal data preprocessing module, a multi-scale feature extraction module, a dynamic adaptive feature selection module based on harmony search, and a dynamic adaptive stacking classification module based on harmony search. The overall process is as follows Figure 1 shown.

[0055] A1 subjects first wore Empatica 4 on their non-dominant wrist and sat quietly with their eyes closed for five minutes to collect physiological baseline signals. Then, the subjects performed cognitive tasks on iPad 2019 and used Apple Pencil to complete drawing-related operations, while their cognitive performance and physiological signals were recorded synchronously throughout the process;

[0056] After obtaining EDA data, IBI data and digital parameters, A2 needs to split and align these data into time series.

[0057] The specific steps of step A2 are as follows:

[0058] A21.EDA time series processing.

[0059] like Figure 2The EDA.csv file downloaded from the cloud is shown, which only contains the session start time and sampling rate information, and lacks the timestamp of each second corresponding to the signal. Therefore, a timestamp is automatically generated for every four data points according to the session start time and the sensor sampling rate (4Hz). Then, the entire EDA time series is divided into baseline data segments and test data segments, with the first 5 minutes of the session as the split point. Finally, the timestamp of the cognitive test is aligned with the EDA series timestamp to extract the EDA signal segment during the corresponding test.

[0060] A22.IBI time series processing.

[0061] During the measurement process of the Empatica 4 wristband, the device may automatically discard the heartbeats that cannot be accurately identified, which causes inconsistencies between the continuous data in the IBI.csv file and the actual distance measurement, such as Figure 3 As shown. In order to ensure the integrity of the data when analyzing the IBI time series in each test task, it is necessary to accurately identify and fill these measurement gaps. To this end, the present invention limits the physiological reasonable range of IBI to within 2 seconds, and data exceeding this threshold will be automatically marked as a measurement gap. Subsequently, these gaps are estimated using cubic spline interpolation to ensure the integrity of the time series. Finally, the time alignment method for IBI data is similar to that for EDA data, including adding new timestamps, dividing into baseline data segments and test data segments, and aligning the timestamps of cognitive tests with the timestamps of the IBI series.

[0062] A3 performs data preprocessing operations on the split and aligned EDA signals, IBI signals and digital cognitive parameters.

[0063] The EDA signal preprocessing operations in step A3 include artifact correction, signal decomposition and overlapping sliding time window processing; the IBI signal preprocessing includes artifact removal, cubic spline interpolation and overlapping sliding time window processing; the digital cognitive parameter preprocessing includes filling in missing values ​​with the mean, removing outliers, and checking data consistency and accuracy.

[0064] The specific steps of step A3 are as follows:

[0065] A31.EDA signal artifact correction.

[0066] like Figure 4 The figure shows the EDA signal preprocessing process. First, the EDA signal is filtered using a first-order Butterworth low-pass filter with a cutoff frequency of 0.6 Hz to retain the low-frequency components in the EDA signal and remove high-frequency noise, so as to more effectively analyze and interpret physiological responses. The transfer function formula of the filter is:

[0067]

[0068] Among them, s is a complex variable, ω c is the cutoff frequency.

[0069] Secondly, the EDAexplorer tool is used to detect artifacts and mark the artifacts of the EDA data after low-pass filtering; then, the artifact signal is interpolated using the cubic spline interpolation method, and the missing data points are approximated by piecewise cubic polynomials to ensure that the function value, first-order derivative, and second-order derivative at each data point are continuous. Suppose there is a set of EDA data (x1, y1), (x2, y2), ..., (x i ,y i ), some of the data points have gaps that need to be interpolated. In each gap interval [x i ,x i+1 ], using a cubic polynomial S i (x) is interpolated:

[0070] S i (x) = a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3

[0071] Among them, a i ,b i ,c i ,d i are the unknown coefficients, which are determined by satisfying the interpolation conditions, including at each data point x i At this point, the function value of the interpolation polynomial is equal to the ordinate value of the data point, that is, S i (x i )=y i ; At each data point x i At , the first-order derivative of the interpolation polynomial is continuous, that is, S i '(x i )=S i ' +1 (x i ); at each data point x i At , the second-order derivative of the interpolation polynomial is continuous, that is, S i ”(x i )=S i ” +1 (x i ).

[0072] A32.EDA signal decomposition.

[0073] EDA is a biological signal that can reflect the physiological and emotional state of an individual, mainly composed of slowly changing basic (Tonic) activities and rapidly changing phase (Phasic) activities. In order to effectively decompose the EDA signal, the present invention adopts the cvxEDA method to decompose the EDA signal into Tonic components, Phasic components and additive white Gaussian noise by solving a convex optimization problem. The optimization problem is restated as a standard quadratic optimization problem, that is, the objective function is a quadratic polynomial of the variables, and the constraints are affine functions, so that it can be solved efficiently under the premise of considering the constraints. By minimizing the loss function, cvxEDA makes the decomposed Phasic and Tonic components fit the original signal as closely as possible, while minimizing the interference of noise. The optimization problem is defined as:

[0074]

[0075] subject to Aq≥0

[0076] Among them, M and A are tridiagonal matrices; q is an auxiliary variable; B is a matrix whose columns are cubic B-spline basis functions; is the spline coefficient vector; C is an N×2 matrix; d is a 2×1 vector containing the offset and slope coefficients of the linear trend; y is a given N sample length EDA signal; parameters α and γ control the penalty intensity for the Tonic and Phasic components respectively. The larger α is, the sparser the estimation result is, and most of the spurious spikes caused by noise will be suppressed, but the signal distortion (that is, the attenuation of true activation) will also be greater. Conversely, the smaller α is, the smaller the distortion is, but the greater the noise is. Regarding γ, the larger the value, the stronger the penalty for , that is, the smoother the Tonic curve.

[0077] After finding the optimal [q, l, d], the Phasic component can be obtained by p = Mq, while the Tonic component can be easily obtained by t = B + Cd. If there is physiological incoherence data, the Tonic and Phasic components can be further filtered using a low-pass Butterworth filter to remove negative SCR and SCL values.

[0078] A33. Overlapping sliding time window processing of EDA signals.

[0079] Considering the minimum time of cognitive test, the present invention sets the window size to 60 seconds, the step size of sliding window to 1 second, and generates a time series according to the first timestamp and the last timestamp of EDA data, which will be used as the starting point of the time window. Assume that the first timestamp is t1 and the last timestamp is t n, the window size is w, and the step size of the sliding window is s, then the value of the time series is [t1, t1+s, t1+2×s, t1+3×s, ..., t n -w].

[0080] A34.IBI signal artifact removal.

[0081] By analyzing the IBI data, heart rate variability (HRV) can be calculated, which indicates the degree of variation between consecutive heart beats. Although the Empatica 4 wristband provides convenience and non-invasiveness for recording HRV, it still faces problems such as artifacts or measurement gaps. Figure 5 Four artifact detection rules are shown to identify IBI artifacts acquired from the Empatica 4 wristband, and the formulas are:

[0082] IBI i ≥250ms IBI i ≤2000ms

[0083] |IBI i -IBI i-1 |≤0.2×IBI i-1

[0084]

[0085]

[0086] A35. Cubic spline interpolation and overlapping sliding time window processing of IBI signals.

[0087] After removing the artifact sequence in the IBI signal, the next step is to apply the cubic spline interpolation method similar to the EDA signal to the IBI signal to fill the missing data. Subsequently, the processed IBI signal is divided into overlapping time windows of 60 seconds in size and 1 second in step size to generate continuous data for subsequent feature extraction.

[0088] A36. Preprocessing of digital cognitive parameters.

[0089] For missing values ​​(such as completion time) caused by data transmission or operational errors, manual filling was performed by manually replaying the video; and for abnormal values ​​(such as handwriting of 0px, score of -1, etc.), these data were directly eliminated because they could not provide effective information for the construction of the MCI diagnostic model. Ultimately, the consistent and accurate digital cognitive parameters after processing will be used for subsequent analysis.

[0090] A4 performs multi-scale feature extraction on EDA data, IBI data and digital parameters under different cognitive tasks.

[0091] The feature extraction in step A4 includes: for different cognitive tasks, constructing feature vectors for EDA signals and HRV signals based on the time domain, constructing feature vectors for EDA signals and HRV signals based on the frequency domain, and constructing feature vectors based on digital cognitive parameters.

[0092] The specific steps of step A4 are as follows:

[0093] A41. The time domain characteristics of EDA signals are the analysis and description of signals in the time dimension. For EDA signals in the time domain, by calculating the SCL (Tonic component) and SCR (Phasic component) of each window, a variety of characteristic indicators are extracted, including mean, standard deviation, median, root mean square, minimum, maximum, energy, skewness, kurtosis and sample entropy. For IBI signals in the time domain, by analyzing the IBI sequence, relevant HRV features are extracted, such as the mean, median, standard deviation, root mean square deviation of the IBI interval, the number of consecutive IBI intervals with a difference of more than 50 milliseconds and the corresponding percentage values, as well as the mean, standard deviation, minimum and maximum values ​​of the heart rate.

[0094] A42. For EDA signal and IBI signal, construct feature vectors based on the frequency domain. First, construct feature vectors for EDA signal based on the frequency domain: use the Welch method to estimate the power spectral density (PSD), and subdivide the Tonic component and Phasic component into five frequency bands within the recommended frequency range (i.e., 0.05-0.50Hz). The Welch method is a frequency domain analysis technique that divides the time series into multiple overlapping windows, and then applies Fourier transform to each window to obtain the PSD estimate of each window. The PSD of each window obtained in this way represents the frequency domain characteristics of the signal in different time periods. Subsequently, the PSD values ​​of all windows are averaged to obtain the overall PSD estimate. On this basis, for each predefined frequency band, the statistical characteristics of its spectral power, such as mean, standard deviation, median, kurtosis, skewness, maximum and minimum values, are calculated.

[0095] Secondly, the HRV signal is decomposed in the frequency domain according to different frequency band ranges, and divided into three frequency bands: high frequency (HF, 0.15Hz to 0.4Hz), low frequency (LF, 0.04Hz to 0.15Hz) and very low frequency (VLF, 0.003Hz to 0.04Hz). At the same time, the Lomb-Scargle method is used to estimate the power spectral density (PSD), and characteristic parameters such as very low frequency power, low frequency power, high frequency power, total power and LF / HF ratio are extracted from it. The formula of the Lomb-Scargle periodogram method is as follows:

[0096]

[0097] Among them, ω j =2πf j , f j is each frequency to be tested, N is the number of data points, t j is the timestamp of each data point, y(t j ) is at time t j The measured value of is the mean value of y(t), σ is the variance of y(t), and τ is the time offset.

[0098] A43. Construct feature vectors based on digitized cognitive parameters: For cognitive tasks, they include the time and score to complete each task, as well as handwriting, frequency, curvature, and sequence features related to the drawing task.

[0099] A5 inputs the extracted multi-scale features into a dynamic adaptive feature selection algorithm based on harmony search to generate the optimal feature subset.

[0100] When processing high-dimensional data, feature selection is particularly important because not all features contribute to the final result, and too many features will lead to dimensionality disaster and model complexity. To this end, the present invention proposes a dynamic adaptive feature selection optimization algorithm based on harmony search, which is used to screen the extracted multi-scale features (time domain, frequency domain, nonlinear domain and cognitive features). The algorithm dynamically adjusts the hyperparameters in each iteration process so that the randomly generated solution vector changes to increase the diversity of the solution, thereby effectively preventing the algorithm from converging to the local optimum. Specifically, in the harmony search method, the Hamming distance is introduced to measure the difference between the currently generated harmony vector and the optimal vector in the harmony memory library. The Hamming distance is used to evaluate the similarity between the two vectors to adjust the search probability. When the Hamming distance is high, the search probability is moderately reduced to enhance the exploratory nature; conversely, the search probability is moderately increased to enhance the tendency to use known information. At the same time, the present invention also optimizes the fitness function value of the harmony vector by using ensemble learning, takes the average performance of all base classifiers as the target, and combines the feature selection rate to improve and balance the classification effect, in order to obtain a feature subset with lower dimension that is suitable for most base classifiers.

[0101] A6 inputs the obtained optimal feature subset into the dynamic adaptive stacking classification model based on harmony search to obtain the final MCI classification result.

[0102] In order to effectively distinguish MCI patients from healthy individuals, the present invention takes efficient and accurate classification as its core goal. Although ensemble learning models usually perform better than single models, simply stacking multiple base learners may cause some learners to perform poorly, and introducing too many base learners will increase the complexity and computational cost of the model. In order to solve this problem, the present invention adopts a harmony search optimization method, and defines the fitness function of the harmony vector as a combination function of the classification error rate and the number of base learners. By iterating based on the selection probability of the harmony memory library, the fitness function value is gradually optimized to minimize it, and finally an accurate MCI detection result is achieved. In addition, the study selected a variety of base learners, including K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Gaussian Naive Bayes (GNB), Support Vector Machine (SVM), Multilayer Perceptron (MLP), Logistic Regression (LR), Gradient Boosting Decision Tree (GBDT) and XGBoost. Logistic Regression (LR) was selected as a meta-learner (or meta-predictor) to integrate the outputs of each base learner to further improve the detection performance.

[0103] Table 1 is a performance comparison of the model proposed in this invention and six machine learning models and the application of the dynamic adaptive feature selection algorithm based on harmony search (AHSI) to model feature selection

[0104]

[0105] As can be seen from Table 1, compared with other algorithms, the model proposed in the present invention shows significant advantages in performance, with an average accuracy of 88.5%. The precision, recall and F1 scores are 89.1%, 88.7% and 88.9% respectively, all at the highest level. It is worth noting that without using AHSI feature selection, SVM has the highest performance among the six evaluated machine learning models, with an accuracy of 79.6%. However, after implementing AHSI feature selection, the performance of AHSI-kNN and AHSI-MLP surpassed AHSI-SVM. The performance of all feature-selected models has been significantly improved, with an average accuracy increase of 3%-5%, and all exceeding 81%. This shows that the AHSI feature selection algorithm has an excellent effect in extracting key features, thereby enhancing the fitting ability of the machine learning model.

[0106] Table 2 is a performance comparison of the model proposed in this invention and five ensemble learning models and the application of the dynamic adaptive feature selection algorithm based on harmony search (AHSI) to model feature selection

[0107]

[0108] RF 0.804±0.015 0.802±0.037 0.816±0.027 0.809±0.025

[0109] GBDT 0.813±0.019 0.816±0.028 0.813±0.027 0.814±0.019

[0110] XGBoost 0.819±0.015 0.825±0.024 0.810±0.026 0.817±0.024

[0111] AHSI-Bag(kNN)0.833±0.024 0.842±0.029 0.821±0.034 0.831±0.036

[0112] AHSI-Bag(SVM)0.811±0.016 0.812±0.024 0.819±0.027 0.815±0.024

[0113] AHSI-RF 0.854±0.017 0.858±0.018 0.847±0.021 0.852±0.024

[0114] AHSI-GBDT 0.848±0.027 0.839±0.025 0.852±0.023 0.845±0.016

[0115] AHSI-XGBoost 0.854±0.022 0.860±0.020 0.849±0.028 0.854±0.029

[0116] Our Method 0.885±0.020 0.891±0.021 0.887±0.024 0.889±0.025

[0117] Consistent with the results in Table 1, the ensemble learning model proposed in the present invention still outperforms other methods in overall performance, especially XGBoost, with an average accuracy of 81.9%, as shown in Table 2. In addition, the bagging model using KNN as a weak learner outperforms the independent KNN model in performance. However, the bagging model with integrated SVM as a weak learner failed to achieve the expected effect, and its performance was even slightly lower than that of the independent SVM model, which may be related to the inherent advantage of SVM in stability.

[0118] Figure 6 (a accuracy, b precision, c recall, dF1 score) are evaluation index comparisons of the model proposed in the present invention and the machine learning model using the dynamic adaptive feature selection algorithm based on harmony search (AHSI).

[0119] As can be seen from the figure, the performance of the model proposed in the present invention is significantly better than other models in all indicators, and the performance is relatively stable without obvious outliers. In contrast, the AHSI-SVM and AHSI-GNB algorithms show more outliers, indicating that it is difficult to accurately fit the data under certain data distributions, resulting in large fluctuations in their performance. Although the AHSI-LR and AHSI-Bag (kNN) algorithms do not have outliers, their upper and lower limits are wide, suggesting that their stability is poor, especially when facing different feature distributions. It is worth noting that the AHSI-RF algorithm shows significant outliers in the F1 score, reflecting the sensitivity of the model to specific data distributions or feature sets.

[0120] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. An MCI detection system based on dynamic adaptive fusion of physiological signals and cognitive data, characterized in that: It includes data acquisition module, time series splitting and alignment module, multimodal data preprocessing module, multi-scale feature extraction module, dynamic adaptive feature selection module based on harmony search, and dynamic adaptive stacking classification module based on harmony search; Data acquisition module: collects physiological baseline signals and performs cognitive tasks, synchronously recording cognitive performance and physiological signals throughout the process; Time series splitting and alignment module: Based on the heterogeneity of the data collected by the data acquisition module, the EDA time series processing strategy and the IBI time series processing strategy are used to process the collected data to maintain the spatiotemporal continuity and effectively deal with potential signal discontinuities, inconsistencies, and misalignments between existing data; Multimodal data preprocessing module: used to process multimodal data that remain consistent in time series, including EDA signal preprocessing, IBI signal preprocessing and digital cognitive parameter preprocessing: The data preprocessing module comprises: EDA signal preprocessing unit: used for EDA signal artifact correction, signal decomposition and overlapping sliding time window processing; IBI signal preprocessing unit: used for IBI signal artifact removal, cubic spline interpolation and overlapping sliding time window processing; Digital cognitive parameter preprocessing unit: used to fill in missing values, remove outliers, and check data consistency and accuracy for the mean of digital cognitive parameters; Multi-scale feature extraction module: for pre-processed multimodal data, feature vectors are constructed for EDA signals and HRV signals based on the time domain, feature vectors are constructed for EDA signals and HRV signals based on the frequency domain, and feature vectors are constructed based on digital cognitive parameters; Dynamic adaptive feature selection module based on harmony search: For the extracted high-dimensional feature set, an improved harmony search algorithm is used to analogize a tone of each instrument to a feature. The hyperparameters are dynamically adjusted in each iteration process so that the randomly generated feature solution vector changes, and finally a feature subset with lower dimension suitable for most base classifiers is obtained; Dynamic adaptive stacked classification module based on harmony search: By improving the harmony search algorithm and making full use of the complementarity and correlation between multimodal data, each basic learner is likened to a tone, and the hyperparameters are dynamically adjusted according to the accuracy and number of learners, ultimately achieving accurate classification of MCI patients.

2. The system according to claim 1, characterized in that The EDA time series processing strategy described is as follows: Since the EDA.csv file downloaded from the cloud only contains the session start time and sampling rate information, and lacks the timestamp corresponding to each signal point, a sampling rate-based timestamp completion strategy is used for processing; a timestamp is automatically generated for every four data points based on the session start time and the sampling rate of 4 Hz; then, the entire EDA time series is divided into a baseline data segment and a test data segment, and the baseline segment is defined as the first 5 minutes after the start of the session; finally, the timestamp of the cognitive task is aligned with the EDA series timestamp to extract the EDA signal segment during the test.

3. The system according to claim 1, characterized in that The IBI time series processing strategy described above: the photoelectric sensor of the Empatica 4 wristband will automatically discard the heartbeat signal that cannot be accurately identified when collecting IBI data, resulting in discontinuity of data in the IBI.csv file; the missing data completion method for the IBI time series is used for processing; first, the physiological reasonable range of IBI is limited to within 2 seconds, and any data beyond this range will be marked as missing; Subsequently, cubic spline interpolation was used to estimate these gaps to ensure the integrity of the time series; the temporal continuity of the IBI data was optimized through curve fitting methods, and finally the IBI data was synchronized with the EDA data by adding timestamps, dividing the baseline and test data segments, and aligning them with the timestamps of the cognitive tasks.

4. The system according to claim 1, characterized in that The EDA signal artifact correction described: First, the EDA signal is filtered using a first-order Butterworth low-pass filter with a cutoff frequency of 0.6 Hz to retain the low-frequency components in the EDA signal while removing high-frequency noise; the transfer function formula of the filter is: Among them, s is a complex variable, ω c is the cutoff frequency; Secondly, the EDAexplorer tool is used to detect artifacts and mark the artifacts of the EDA data after low-pass filtering. Then, the artifact signal is interpolated using the cubic spline interpolation method, and the missing data points are approximated by piecewise cubic polynomials to ensure that the function value, first-order derivative, and second-order derivative at each data point are continuous. Suppose there is a set of EDA data (x1, y1), (x2, y2), ..., (x i ,y i ), some of the data points have gaps that need to be interpolated; in each gap interval [x i ,x i+1 ], using a cubic polynomial S i (x) is interpolated: S i (x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 Among them, a i ,b i ,c i ,d i are the unknown coefficients, which are determined by satisfying the interpolation conditions, including at each data point x i At this point, the function value of the interpolation polynomial is equal to the ordinate value of the data point, that is, S i (x i )=y i ; At each data point x i At, the first-order derivative of the interpolation polynomial is continuous, that is, S′ i (x i ) = S′ i+1 (x i ); at each data point x i At , the second-order derivative of the interpolation polynomial is continuous, that is, S i ”(x i )=S i ” +1 (x i ).

5. The system according to claim 1, characterized in that The EDA signal decomposition described: cvxEDA regards the EDA signal as a superposition of Tonic components, Phasic components and additive white Gaussian noise by solving a convex optimization problem; the optimization problem is restated as a standard quadratic optimization, that is, the objective function is a quadratic polynomial of the variables, and the constraints are affine functions, so that efficient solution is achieved while considering the constraints; by minimizing the loss function, cvxEDA makes the decomposed Phasic and Tonic components fit the original signal as much as possible and minimizes the influence of noise; This optimization problem is defined as: subject to Aq≥0 Among them, M and A are both tridiagonal matrices; q is an auxiliary variable; B is a matrix whose columns are cubic B-spline basis functions; l is the spline coefficient vector; C is an N×2 matrix; d is a 2×1 vector containing the offset and slope coefficients of the linear trend; y is a given N sample length EDA signal; parameters α and γ control the penalty intensity for the Tonic and Phasic components respectively; the larger α is, the sparser the estimation result is, and most of the spurious spikes caused by noise will be suppressed, but the signal distortion will also be greater; conversely, the smaller α is, the smaller the distortion is, but the noise is greater; regarding γ, the larger the value is, the stronger the penalty for l is, that is, the smoother the Tonic curve is; After finding the optimal [q, l, d], the Phasic component can be obtained by p = Mq, and the Tonic component can be easily obtained by t = Bl + Cd; If physiological incoherent data are present, the Tonic and Phasic components can be further filtered using a low-pass Butterworth filter to remove negative SCR and SCL values.

6. The system according to claim 1, characterized in that The EDA signal overlaps the sliding time window processing: considering the minimum time of cognitive test, the window size is set to 60 seconds, the step size of the sliding window is 1 second, and a time series is generated according to the first and last timestamps of the EDA data, which will be used as the starting point of the time window; Assume the first timestamp is t1 and the last timestamp is t n , the window size is w, and the step size of the sliding window is s, then the value of the time series is [t1, t1+s, t1+2×s, t1+3×s, ..., t n -w].

7. The system according to claim 1, characterized in that The IBI signal artifact removal: Four artifact detection rules are used to identify IBI artifacts obtained from the Empatica 4 wristband, and their formulas are: PLACE i ≥250ms∩IBI i ≤2000ms |LOCATION i -IBI i-1 |≤0.2×IBI i-1 8. The system according to claim 1, characterized in that The feature vectors of EDA signals and HRV signals are constructed based on the time domain: the mean, standard deviation, median, root mean square, minimum, maximum, energy, skewness, kurtosis, sample entropy and other EDA time domain features are extracted by calculating the SCL (Tonic component) and SCR (Phasic component) of each window; mean, median, standard deviation, root mean square deviation of IBI intervals, the number of consecutive IBI intervals with a difference of more than 50 milliseconds and the corresponding percentage values, and mean, standard deviation, minimum and maximum values ​​of heart rate and other HRV time domain features are extracted by calculating the IBI sequence.

9. The system according to claim 1, characterized in that The described method of constructing a feature vector for an EDA signal based on the frequency domain: using the Welch method to estimate the power spectral density, and subdividing the Tonic component and the Phasic component into five frequency bands within the recommended frequency range; the Welch method is to divide the time series into multiple overlapping windows, and then apply Fourier transform to each window to obtain the PSD estimation of each window; the obtained PSD of each window represents the frequency domain characteristics of the signal in different time periods; then, the PSD values ​​of all windows are averaged to obtain the overall PSD estimation; on this basis, for each predefined frequency band, the statistical characteristics of its spectral power are calculated, including the mean, standard deviation, median, kurtosis, skewness, maximum value and minimum value; According to different frequency band ranges, the HRV signal is decomposed in the frequency domain and divided into three frequency bands: high frequency (HF, 0.15Hz to 0.4Hz), low frequency (LF, 0.04Hz to 0.15Hz) and very low frequency (VLF, 0.003Hz to 0.04Hz). At the same time, the Lomb-Scargle method is used to estimate the power spectrum density, and characteristic parameters such as very low frequency power, low frequency power, high frequency power, total power and LF / HF ratio are extracted from it. The formula of the Lomb-Scargle periodogram method is as follows: Among them, ω j =2πf j , f j is each frequency to be tested, N is the number of data points, t j is the timestamp of each data point, y(t j ) is at time t j The measured value of is the mean value of y(t), σ is the variance of y(t), and τ is the time offset.

10. The system according to claim 1, characterized in that For dynamic adaptive feature selection based on harmony search, the Hamming distance is introduced to measure the difference between the currently generated harmony vector and the optimal vector in the harmony memory library; the Hamming distance is used to evaluate the similarity between the two vectors to adjust the search probability; when the Hamming distance is high, the search probability is moderately reduced to enhance the exploratory nature; conversely, the search probability is moderately increased to enhance the tendency to utilize known information; the fitness function value of the harmony vector is also optimized using ensemble learning, with the average performance of all base classifiers as the goal, and the feature selection rate is combined to improve and balance the classification effect, in order to obtain a feature subset with lower dimensionality that is suitable for most base classifiers.

Images