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

Abstract

The application discloses a kind of MCI detection systems based on physiological signal and dynamic adaptive fusion of cognitive data, including data acquisition module, time series split and alignment module, multi-modal data preprocessing module, multiscale feature extraction module, dynamic adaptive feature selection module based on harmony search, dynamic adaptive stacking classification module based on harmony search;The application proposes dynamic adaptive feature selection optimization algorithm based on harmony search, reduces redundant information and optimizes feature set.Finally, by proposing dynamic adaptive stacking classification optimization algorithm based on harmony search, the accurate classification of MCI patients and healthy individuals is realized.The application combines conventional wearable devices with cognitive assessment on tablet computer, with the advantages of simple operation, suitable for home environment, etc., providing important technical support for self-detection MCI.

Description

Technical Field

[0001] This invention relates to the field of mild cognitive impairment detection technology in medical diagnosis, specifically an MCI detection system based on dynamic adaptive fusion of physiological signals and cognitive data. Background Technology

[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 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 are currently no drugs that can completely cure dementia or significantly alter its clinical course. Furthermore, research indicates that early intervention is most effective, which requires accurate and sensitive diagnosis of MCI. Therefore, effective MCI detection is crucial, as it may enable patients to receive timely intervention, thereby slowing cognitive decline and alleviating the burden of dementia.

[0003] Currently, abundant individual physiological data collected through wearable devices provides a new avenue for detecting the physiological and behavioral characteristics of patients with cognitive impairment, enabling timely, economical, and efficient detection of mild cognitive impairment (MCI). Studies have shown that physiological indicators such as heart rate variability, skin conductance, gait variability, skin temperature, respiratory rate, electroencephalography (EEG), eye tracking, and electromyography (EMG) can effectively reflect changes in cognitive function, providing objective evidence for the early diagnosis of cognitive impairment. However, due to the diversity and individual variability of these physiological indicators, selecting appropriate monitoring parameters becomes particularly complex, increasing the difficulty for doctors and patients in making decisions using physiological monitoring data. Furthermore, existing cognitive function monitoring practices mainly focus on the analysis of physiological parameters, lacking comprehensive analysis of these physiological data with the results of traditional or computerized cognitive tests. This limits the comprehensive assessment of cognitive function and fails to fully leverage the complementary potential of physiological data and cognitive test results. Therefore, how to efficiently integrate and evaluate diverse physiological measurement data, and adaptively fuse indicators with significant monitoring value with cognitive test measurement parameters to achieve comprehensive cognitive function assessment, has become an urgent problem to be solved. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a dynamic adaptive fusion system for MCI detection based on physiological signals and cognitive data, thereby improving the accuracy and practicality of MCI detection. In this system, firstly, time-series alignment of electrodermal activity, photoplethysmography (PPG), and digital cognitive parameters is performed using cubic spline interpolation and timestamp sorting matching techniques. Secondly, a multi-stage automatic artifact removal algorithm is designed, combining artifact correction, signal decomposition, and overlapping sliding time window techniques for data purification. Based on this, multi-scale features are extracted using time-domain, frequency-domain, and nonlinear analysis methods. Then, a dynamic adaptive feature selection algorithm based on harmony search is proposed, which adaptively updates the harmony memory to select the most helpful feature subset for classification, while introducing adaptive hyperparameter adjustment to iteratively improve classification performance. This invention, by innovatively combining physiological data and cognitive tests, provides a more accurate and practical detection method for the early diagnosis of MCI.

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

[0006] A dynamic adaptive fusion MCI detection system based on physiological signals and cognitive data is 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 acquisition module: Participants first wore the Empatica 4 on their non-dominant wrist and sat quietly with their eyes closed for five minutes to collect baseline physiological signals. Then, participants performed cognitive tasks on an iPad 2019, using Apple Pencil to complete drawing-related operations, with their cognitive performance and physiological signals recorded synchronously throughout the process.

[0008] Time series splitting and alignment module: Based on the heterogeneity of collected data (such as EDA, IBI and cognitive parameters), this 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 data.

[0009] Multimodal data preprocessing module: Used to process multimodal data that remains consistent over time, including EDA signal preprocessing, IBI signal preprocessing, and digital cognitive parameter preprocessing.

[0010] The data preprocessing module includes:

[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 ​​for the mean of digital cognitive parameters, remove outliers, and check data consistency and accuracy;

[0014] Multi-scale feature extraction module: For preprocessed multimodal data, this invention constructs feature vectors based on EDA signals and HRV signals in the time domain, feature vectors based on EDA signals and HRV signals in the frequency domain, and feature vectors based on digital cognitive parameters.

[0015] The 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, which compares a pitch of each instrument to a feature. In each iteration, the hyperparameters are dynamically adjusted so that the randomly generated feature solution vector changes, and finally a feature subset with low dimension that is applicable to most base classifiers is obtained.

[0016] The dynamic adaptive stacked classification module based on harmony search: This 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 hyperparameters according to the accuracy and number of learners, ultimately achieving accurate classification of MCI patients.

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

[0018] The system described above employs the following IBI time series processing strategy: The photoelectric sensor on the Empatica 4 wristband automatically discards heartbeat signals that cannot be accurately identified when collecting IBI data, leading to discontinuities in the IBI.csv file. Therefore, this invention utilizes a method for missing data completion in the IBI time series. First, the physiologically reasonable range of IBI is limited to within 2 seconds; any data exceeding this range is marked as missing. Subsequently, cubic spline interpolation is used to estimate these gaps to ensure the integrity of the time series. Curve fitting is used to optimize the temporal continuity of the IBI data. Finally, the IBI data is processed synchronously with EDA data, timestamps are added, baselines and test data segments are defined, and the data is aligned with the timestamps of the cognitive task.

[0019] The system described above, specifically the EDA signal artifact correction, involves: first, using a first-order Butterworth low-pass filter with a cutoff frequency of 0.6Hz to filter the EDA signal, preserving the low-frequency components while removing high-frequency noise; the transfer function formula for this filter is:

[0020]

[0021] Where s is a complex variable, ω c It is the cutoff frequency.

[0022] Secondly, artifact detection is performed using the EDAexplorer tool to mark artifacts in the low-pass filtered EDA data. Then, cubic spline interpolation is used to interpolate the artifact signal, approximating the missing data points using piecewise cubic polynomials to ensure the continuity of the function value, first derivative, and second derivative at each data point. Assume there is a set of EDA data (x1, y1), (x2, y2), ..., (x... i y i ), where some data points have gaps that need to be interpolated. In each gap interval [x i ,x i+1 Within, a cubic polynomial S is used. 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 These are undetermined coefficients, determined by satisfying interpolation conditions, including at each data point x. i At point S, the function value of the interpolation polynomial is equal to the ordinate value of the data point, i.e., S i (x i )=y i ; at each data point x i At point S, the first derivative of the interpolation polynomial is continuous, i.e., S i '(x i ) = S i ' +1 (x i ); at each data point x i At point S, the second derivative of the interpolation polynomial is continuous, i.e., S i (x) i ) = S i " +1 (x i ).

[0025] The system described above, specifically the EDA signal decomposition: cvxEDA solves a convex optimization problem, treating the EDA signal as a superposition of Tonic, Phasic, and additive white Gaussian noise. The optimization problem is reformulated as a standard quadratic optimization (i.e., the objective function is a quadratic polynomial of variables, and the constraints are affine functions), thus achieving efficient solution while considering the constraints. By minimizing the loss function, cvxEDA ensures that the decomposed Phasic and Tonic components fit the original signal as closely as possible and minimizes the influence of noise. This optimization problem is defined as:

[0026]

[0027] subject to Aq≥0

[0028] Where 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 EDA signal of N sample lengths; parameters α and γ control the penalty strength for the Tonic and Phasic components, respectively. A larger α results in a sparser estimation, suppressing most spurious spikes caused by noise, but also increases signal distortion (i.e., attenuation of true activation). Conversely, a smaller α results in less distortion but more noise. Regarding γ, a larger value means a stronger penalty, resulting in a smoother Tonic curve.

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

[0030] The system described above, specifically the EDA signal overlapping sliding time window processing, involves the following: Given the minimum time requirement for cognitive testing, the window size is set to 60 seconds, and the sliding window step size is 1 second. Simultaneously, a time series is generated based on the first and last timestamps of the EDA data; this series serves as the starting point of the time window. Assume the first timestamp is t1 and the last timestamp is t... n Given a window size of w and a sliding window step size of s, the value of this time series is [t1, t1+s, t1+2×s, t1+3×s, ..., t n -w].

[0031] The system described above, specifically the IBI signal artifact removal, employs four artifact detection rules to identify IBI artifacts acquired from the Empatica 4 wristband. The formulas for these rules are as follows:

[0032] IBI i ≥250msIBI i ≤2000ms

[0033] |IBI i -IBI i-1 |≤0.2×IBI i-1

[0034]

[0035]

[0036] The system described above includes cubic spline interpolation and overlapping sliding time window processing of the IBI signal: applying a cubic spline interpolation method similar to that used for EDA signals to fill in missing values ​​to the detected IBI artifact signals; and dividing the processed IBI signal into overlapping time windows of 60 seconds in size and 1 second in step size to generate continuous data for subsequent feature extraction.

[0037] The system described above constructs feature vectors based on time-domain EDA and HRV signals by: extracting EDA time-domain features such as mean, standard deviation, median, root mean square, minimum, maximum, energy, skewness, kurtosis, and sample entropy by calculating the SCL (Tonic component) and SCR (Phasic component) of each window; and extracting HRV time-domain features such as mean, median, standard deviation, root mean square difference, number of consecutive IBI intervals with a difference of more than 50 milliseconds and corresponding percentage values ​​by calculating the IBI sequence; and extracting HRV time-domain features such as mean, standard deviation, minimum, and maximum heart rate.

[0038] The system described above constructs feature vectors for EDA signals based on the frequency domain by employing the Welch method for power spectral density (PSD) estimation and subdividing the tonic and phasic components into five frequency bands within a recommended frequency range (0.05-0.50 Hz). The Welch method is a frequency domain analysis technique that divides the time series into multiple overlapping windows and then applies a Fourier transform to each window to obtain the PSD estimate for each window. The PSD of each window thus 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. Based on this, for each predefined frequency band, statistical characteristics of its spectral power are calculated, such as mean, standard deviation, median, kurtosis, skewness, maximum, and minimum values.

[0039] Furthermore, the HRV signal was frequency-domain decomposed according to different frequency bands, dividing it into three bands: high frequency (HF, 0.15Hz to 0.4Hz), low frequency (LF, 0.04Hz to 0.15Hz), and extremely low frequency (VLF, 0.003Hz to 0.04Hz). Simultaneously, the Lomb-Scargle method was used to estimate the power spectral density (PSD), and characteristic parameters such as extremely low frequency power, low frequency power, high frequency power, total power, and the LF / HF ratio were extracted. The formula for the Lomb-Scargle periodogram method is as follows:

[0040]

[0041] Where, ω j =2πf j f j Here, t represents the frequency of each test point, N is the number of data points, and t is the frequency of each test point. j It is the timestamp of each data point, y(t) j ) is at time t j The measured value, σ is the average value of y(t), σ is the variance of y(t), and τ is the time offset.

[0042] The system, which constructs feature vectors based on digital cognitive parameters, includes, for cognitive tasks, 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 described herein, for dynamic adaptive feature selection based on harmony search, introduces Hamming distance to measure the difference between the currently generated harmony vector and the optimal vector in the harmony memory. 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 exploration; conversely, the search probability is moderately increased to enhance the tendency to utilize known information. Simultaneously, this invention also utilizes ensemble learning to optimize the fitness function value of the harmony vector, using the average performance of all base classifiers as the target, and combining it with the feature selection rate to improve and balance classification performance, aiming to obtain a low-dimensional feature subset applicable to most base classifiers.

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

[0045] The beneficial effects of this invention are:

[0046] This invention constructs a dynamic adaptive ensemble 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. To address the feature redundancy problem, this invention introduces an improved harmony search feature selection algorithm. This algorithm effectively avoids the trap of local optima through an adaptive hyperparameter adjustment strategy and establishes a comprehensive evaluation index to extract a more comprehensive feature subset on each base learner. Subsequently, the system employs nine heterogeneous learners as base 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 this invention are characterized by high cost-effectiveness, excellent diagnostic accuracy, and good transferability. Its simplicity and ease of implementation make it demonstrate broad application potential in the clinical auxiliary diagnosis of neurocognitive disorders, thus providing the medical field with more accurate and personalized diagnostic and treatment solutions. Attached Figure Description

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

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

[0049] Figure 3 This is a schematic diagram illustrating the splitting and time alignment process of IBI data according to 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-based signal preprocessing workflow of the present invention;

[0052] Figure 6 This diagram illustrates a comparison of evaluation metrics between the model proposed in this invention and a machine learning model that applies the Dynamic Adaptive Feature Selection Algorithm (AHSI) based on harmony search. Detailed Implementation

[0053] The present invention will be described in detail below with reference to specific embodiments.

[0054] A dynamic adaptive fusion system for MCI detection based on physiological signals and cognitive data is 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 As shown.

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

[0056] After obtaining EDA data, IBI data, and digitization parameters, A2 needs to perform time series splitting and alignment on these data.

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

[0058] A21. EDA Time Series Processing.

[0059] like Figure 2The image shows an EDA.csv file downloaded from the cloud, which only contains session start time and sampling rate information, lacking a timestamp per second corresponding to the signal. Therefore, based on the session start time and sensor sampling rate (4Hz), a timestamp is automatically generated for every four data points. The entire EDA time series is then divided into baseline and test data segments, with the first 5 minutes of the session as the dividing point. Finally, the timestamps from the cognitive test are aligned with the EDA sequence timestamps to extract the corresponding EDA signal segments from the test process.

[0060] A22. IBI Time Series Processing.

[0061] During measurements with the Empatica 4 wristband, inaccurately identified heartbeats may be automatically discarded by the device, leading to inconsistencies between the continuous data in the IBI.csv file and the actual distance measurements. Figure 3 As shown. To ensure data integrity when analyzing IBI time series in various test tasks, it is necessary to accurately identify and fill these measurement gaps. Therefore, this invention limits the physiologically reasonable range of IBI to within 2 seconds; data exceeding this threshold are automatically marked as measurement gaps. Subsequently, cubic spline interpolation is used to estimate these gaps 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 the data into baseline and test data segments, and aligning the timestamps of the cognitive tests with the timestamps of the IBI sequences.

[0062] A3 performs data preprocessing operations on the split and aligned EDA signals, IBI signals, and digitized 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 preprocessing of digital cognitive parameters 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 diagram shows the EDA signal preprocessing flow. First, a first-order Butterworth low-pass filter with a cutoff frequency of 0.6Hz is used to filter the EDA signal, preserving the low-frequency components while removing high-frequency noise, thus enabling more effective analysis and interpretation of physiological responses. The transfer function formula for this filter is:

[0067]

[0068] Where s is a complex variable, ω c It is the cutoff frequency.

[0069] Secondly, artifact detection is performed using the EDAexplorer tool to mark artifacts in the low-pass filtered EDA data. Then, cubic spline interpolation is used to interpolate the artifact signal, approximating the missing data points using piecewise cubic polynomials to ensure the continuity of the function value, first derivative, and second derivative at each data point. Assume there is a set of EDA data (x1, y1), (x2, y2), ..., (x... i y i ), where some data points have gaps that need to be interpolated. In each gap interval [x i ,x i+1 Within, a cubic polynomial S is used. 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 These are undetermined coefficients, determined by satisfying interpolation conditions, including at each data point x. i At point S, the function value of the interpolation polynomial is equal to the ordinate value of the data point, i.e., S i (x i )=y i ; at each data point x i At point S, the first derivative of the interpolation polynomial is continuous, i.e., S i '(x i ) = S i ' +1 (x i ); at each data point x i At point S, the second derivative of the interpolation polynomial is continuous, i.e., S i (x) i ) = S i " +1 (x i ).

[0072] A32. EDA signal decomposition.

[0073] EDA (Electrode Analytic Acid) is a biosignal that reflects an individual's physiological and emotional state, primarily composed of slowly changing tonic activity and rapidly changing phase activity. To effectively decompose EDA signals, this invention employs the cvxEDA method. By solving a convex optimization problem, the EDA signal is decomposed into tonic components, phase components, and additive white Gaussian noise. This optimization problem is reformulated as a standard quadratic optimization problem, i.e., a quadratic polynomial with a variable objective function and affine constraints, enabling efficient solution while considering the constraints. By minimizing the loss function, cvxEDA ensures that the decomposed phase and tonic components closely resemble the original signal while minimizing noise interference. The optimization problem is defined as follows:

[0074]

[0075] subject to Aq≥0

[0076] Where 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 EDA signal of N sample lengths; parameters α and γ control the penalty strength for the Tonic and Phasic components, respectively. A larger α results in a sparser estimation, suppressing most spurious spikes caused by noise, but also increases signal distortion (i.e., attenuation of true activation). Conversely, a smaller α results in less distortion but more noise. Regarding γ, a larger value means a stronger penalty, resulting in a smoother Tonic curve.

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

[0078] A33. EDA signal overlap sliding time window processing.

[0079] Considering the minimum time required for cognitive testing, this invention sets the window size to 60 seconds and the sliding window step size to 1 second. Simultaneously, a time series is generated based on the first and last timestamps of the EDA data; this series will serve as the starting point of the time window. Assume the first timestamp is t1 and the last timestamp is t... nGiven a window size of w and a sliding window step size of s, the value of this time series is [t1, t1+s, t1+2×s, t1+3×s, ..., t n -w].

[0080] A34. IBI signal artifact removal.

[0081] Analysis of IBI data allows for the calculation of Heart Rate Variability (HRV), which represents the degree of variation between consecutive heartbeats. While the Empatica 4 wristband offers convenience and non-invasiveness for HRV recording, it still faces issues such as artifacts or measurement gaps. Figure 5 The diagram shows four artifact detection rules used to identify IBI artifacts acquired from the Empatica 4 wristband. The formulas are as follows:

[0082] IBI i ≥250msIBI i ≤2000ms

[0083] |IBI i -IBI i-1 |≤0.2×IBI i-1

[0084]

[0085]

[0086] A35. IBI signal cubic spline interpolation and overlapping sliding time window processing.

[0087] After removing artifact sequences from the IBI signal, the next step is to apply a cubic spline interpolation method similar to that used for EDA signals to fill in missing data. Subsequently, the processed IBI signal is divided into overlapping time windows of 60 seconds each with a step size of 1 second to generate continuous data, facilitating 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 is performed by replaying the video. Outliers (such as handwriting of 0px or a score of -1) are discarded because they do not provide valid information for building the MCI diagnostic model. Finally, the processed, consistent, and accurate digital cognitive parameters will be used for subsequent analysis.

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

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

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

[0093] A41. The time-domain characteristics of EDA signals refer to the analysis and description of the signal in the time dimension. For EDA signals in the time domain, various feature indicators are extracted by calculating the SCL (Tonic component) and SCR (Phasic component) of each window, including mean, standard deviation, median, root mean square, minimum, maximum, energy, skewness, kurtosis, and sample entropy. For IBI signals in the time domain, relevant HRV features are extracted by analyzing the IBI sequence, such as the mean, median, standard deviation, root mean square difference 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 heart rate.

[0094] A42. For EDA and IBI signals, feature vectors are constructed based on the frequency domain. First, feature vectors are constructed for the EDA signal based on the frequency domain: the Welch method is used to estimate the power spectral density (PSD), and the tonic and phasic components are subdivided into five frequency bands within a 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 a Fourier transform to each window to obtain the PSD estimate for each window. The PSD of each window thus 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. Based on this, for each predefined frequency band, the statistical characteristics of its spectral power are calculated, such as mean, standard deviation, median, kurtosis, skewness, maximum, and minimum values.

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

[0096]

[0097] Where, ω j =2πf j f j Here, t represents the frequency of each test point, N is the number of data points, and t is the frequency of each test point. j It is the timestamp of each data point, y(t) j ) is at time t j The measured value, σ is the average value of y(t), σ is the variance of y(t), and τ is the time offset.

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

[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] Feature selection is crucial when processing high-dimensional data, as not all features contribute to the final result, and an excessive number of features can lead to the curse of dimensionality and model complexity. To address this, this invention proposes a dynamic adaptive feature selection optimization algorithm based on harmony search to filter extracted multi-scale features (time domain, frequency domain, nonlinear domain, and cognitive features). This algorithm dynamically adjusts hyperparameters during each iteration, changing the randomly generated solution vectors to increase solution diversity and effectively prevent convergence to local optima. Specifically, in the harmony search method, Hamming distance is introduced to measure the difference between the currently generated harmony vector and the optimal vector in the harmony memory. 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 exploration; conversely, the search probability is moderately increased to enhance the tendency to utilize known information. Simultaneously, this invention also utilizes ensemble learning to optimize the fitness function value of the harmony vectors, using the average performance of all base classifiers as the target, and combining it with the feature selection rate to improve and balance classification performance, aiming to obtain a feature subset with low dimensionality applicable to most base classifiers.

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

[0102] To effectively distinguish between MCI patients and healthy individuals, this invention aims to achieve efficient and accurate classification. While ensemble learning models generally outperform single models, simply stacking multiple base learners can lead to poor performance from some learners, and introducing too many base learners increases model complexity and computational cost. To address this issue, this invention employs a harmony search optimization method, defining the fitness function of the harmony vector as a combination of the classification error rate and the number of base learners. By iteratively optimizing the fitness function value based on the selection probability of the harmony memory library, it ultimately achieves accurate MCI detection results. Furthermore, the study selected various 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 chosen as the meta-learner (or meta-predictor) to integrate the outputs of the various base learners, further improving detection performance.

[0103] Table 1 compares the performance of the model proposed in this invention with six machine learning models, as well as the performance of applying the Dynamic Adaptive Feature Selection Algorithm (AHSI) based on harmony search to the model's feature selection.

[0104]

[0105] As shown in Table 1, compared with other algorithms, the model proposed in this invention exhibits a significant performance advantage, achieving an average accuracy of 88.5%. Precision, recall, and F1 score are 89.1%, 88.7%, and 88.9%, respectively, all at the highest levels. Notably, without AHSI feature selection, SVM performs best among the six evaluated machine learning models, achieving an accuracy of 79.6%. However, after implementing AHSI feature selection, AHSI-kNN and AHSI-MLP outperform AHSI-SVM. All models with feature selection show significant performance improvements, with average accuracy increasing by 3%-5%, all exceeding 81%. This indicates that the AHSI feature selection algorithm has excellent effects in extracting key features, thereby enhancing the fitting ability of machine learning models.

[0106] Table 2 compares the performance of the proposed model, five ensemble learning models, and the application of the Harmony Search-based Dynamic Adaptive Feature Selection (AHSI) algorithm to model feature selection.

[0107]

[0108]

[0109] Consistent with the results in Table 1, the ensemble learning model proposed in this invention still outperforms other methods in overall performance, especially XGBoost, which achieves an average accuracy of 81.9%, as shown in Table 2. Furthermore, the bagging model using KNN as a weak learner outperforms the standalone KNN model. However, the bagging model integrating SVM as a weak learner failed to achieve the expected results, and its performance was even slightly lower than the standalone SVM model. This may be related to the inherent stability advantage of SVM.

[0110] Figure 6 (a. Accuracy, b. Precision, c. Recall, dF1 score) are evaluation metrics comparing the model proposed in this invention with a machine learning model applying the Dynamic Adaptive Feature Selection Algorithm (AHSI) based on harmony search.

[0111] As shown in the figure, the model proposed in this invention significantly outperforms other models across all metrics and exhibits relatively stable performance without significant outliers. In contrast, the AHSI-SVM and AHSI-GNB algorithms show more outliers, indicating difficulty in accurately fitting data under certain distributions, leading to significant performance fluctuations. Although the AHSI-LR and AHSI-Bag(kNN) algorithms do not show outliers, their wide upper and lower bounds suggest poor stability, especially when facing different feature distributions. Notably, the AHSI-RF algorithm exhibits significant outliers in its F1 score, reflecting the model's sensitivity to specific data distributions or feature sets.

[0112] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An MCI detection system based on dynamic adaptive fusion of physiological signals and cognitive data, characterized in that, It includes 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. Data acquisition module: Acquires 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 module uses EDA time series processing strategy and IBI time series processing strategy to process the collected data in order to maintain spatiotemporal continuity and effectively deal with potential signal discontinuity, inconsistency and misalignment between data. Multimodal data preprocessing module: Used to process multimodal data that remains consistent over time, including EDA signal preprocessing, IBI signal preprocessing, and digital cognitive parameter preprocessing. The data preprocessing module includes: 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 ​​for the mean of digital cognitive parameters, remove outliers, and check data consistency and accuracy; Multi-scale feature extraction module: For preprocessed multimodal data, feature vectors are constructed based on EDA signals and HRV signals in the time domain, feature vectors are constructed based on EDA signals and HRV signals in the frequency domain, and feature vectors are constructed based on digital cognitive parameters. The dynamic adaptive feature selection module based on harmony search: For the extracted high-dimensional feature set, an improved harmony search algorithm is adopted, which treats a pitch of each instrument as a feature. The hyperparameters are dynamically adjusted in each iteration, so that the randomly generated feature solution vector changes, and finally a feature subset with low dimension that is suitable for most base classifiers is obtained. To address dynamic adaptive feature selection based on harmony search, Hamming distance is introduced to measure the difference between the currently generated harmony vector and the optimal vector in the harmony memory. 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 exploration; conversely, the search probability is moderately increased to enhance the tendency to utilize known information. Ensemble learning is also used to optimize the fitness function value of the harmony vector, with the average performance of all base classifiers as the target, and the feature selection rate is combined to improve and balance the classification effect, aiming to obtain a feature subset with low dimensionality that is applicable to most base classifiers. The dynamic adaptive stacked classification module based on harmony search: By improving the harmony search algorithm, it makes full use of the complementarity and correlation between multimodal data, analogizes each basic learner to a tone, and dynamically adjusts hyperparameters 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 is as follows: Since the EDA.csv file downloaded from the cloud only contains session start time and sampling rate information, and lacks timestamps corresponding to each signal point, a timestamp completion strategy based on sampling rate is adopted for processing; 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, with the baseline segment defined as the first 5 minutes after the session starts; finally, the timestamps of the cognitive task are aligned with the timestamps of the EDA sequence to extract the EDA signal segment during the test period.

3. The system according to claim 1, characterized in that, The IBI time series processing strategy is as follows: When the photoelectric sensor of the Empatica4 wristband collects IBI data, it will automatically discard heartbeat signals that cannot be accurately identified, resulting in discontinuity of data in the IBI.csv file; a missing data completion method for IBI time series is adopted for processing; firstly, the physiologically reasonable range of IBI is limited to within 2 seconds, and any data outside 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 by curve fitting. Finally, the IBI data and EDA data were synchronized, timestamps were added, baselines and test data segments were defined, and the timestamps were aligned with those of the cognitive task.

4. The system according to claim 1, characterized in that, The EDA signal artifact correction is as follows: First, a first-order Butterworth low-pass filter with a cutoff frequency of 0.6Hz is used to filter the EDA signal, preserving the low-frequency components while removing high-frequency noise; the transfer function formula of this filter is: ;in, It is a complex variable. It is the cutoff frequency; Secondly, artifact detection is performed using the EDAexplorer tool to mark artifacts in the low-pass filtered EDA data. Then, cubic spline interpolation is used to interpolate the artifact signal, approximating the missing data points using piecewise cubic polynomials to ensure the continuity of the function value, first derivative, and second derivative at each data point. Assume there is a set of EDA data... Some data points have gaps that need to be interpolated; in each gap interval Inside, use a cubic polynomial Perform interpolation: ； in, These are undetermined coefficients, determined by satisfying interpolation conditions, including at each data point. At this point, the function value of the interpolation polynomial is equal to the ordinate value of the data point, that is... ; at each data point At this point, the first derivative of the interpolation polynomial is continuous, i.e. ; at each data point At this point, the second derivative of the interpolation polynomial is continuous, i.e. .

5. The system according to claim 1, characterized in that, The EDA signal decomposition described above: cvxEDA solves a convex optimization problem, treating the EDA signal as a superposition of Tonic components, Phasic components, and additive white Gaussian noise. The optimization problem is reformulated as a standard quadratic optimization, where the objective function is a quadratic polynomial of the variables and the constraints are affine functions, thus achieving efficient solution while considering the constraints. By minimizing the loss function, cvxEDA makes the decomposed Phasic and Tonic components fit the original signal as closely as possible and minimizes the influence of noise. This optimization problem is defined as: ; in, and All are tridiagonal matrices; As an auxiliary variable; It is a matrix whose columns are cubic. spline basis functions; It is a vector of spline coefficients; It is an N×2 matrix; It is a 2×1 vector containing the offset and slope coefficient of the linear trend; Given an EDA signal with a length of N samples; parameters and Control the penalty intensity for the Tonic and Phasic components separately; The larger the value, the sparser the estimation results, and most of the spurious spikes caused by noise will be suppressed, but the signal distortion will also be greater; conversely, The smaller the value, the less distortion, but the greater the noise; regarding The larger the value, the better. The stronger the penalty, the smoother the Tonic curve; Finding the optimal Afterwards, the Phasic component can be derived from... It is concluded that the Tonic component can be obtained through Easily obtained; If there is physiological incoherence, a low-pass Butterworth filter can be used to further filter the tonic and phasic components to remove negative SCR and SCL values.

6. The system according to claim 1, characterized in that, The EDA signal overlapping sliding time window processing is as follows: In view of the minimum time of the 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 based on the first and last timestamps of the EDA data. This series will be used as the starting point of the time window. Assuming the first timestamp is The last timestamp is Window size is The step size of the sliding window is Then the value of this time series is .

7. The system according to claim 1, characterized in that, The IBI signal artifact removal described above employs four artifact detection rules to identify IBI artifacts acquired from the Empatica 4 wristband, with the following formulas: ; ; ; .

8. The system according to claim 1, characterized in that, The aforementioned construction of feature vectors based on time-domain EDA and HRV signals involves: calculating the SCL and SCR for each window, with the SCL using the Tonic component and the SCR using the Phasic component, and extracting the following EDA time-domain features: mean, standard deviation, median, root mean square, minimum, maximum, energy, skewness, kurtosis, and sample entropy. Simultaneously, by calculating the IBI sequence, the following HRV time-domain features are extracted: mean, median, standard deviation, root mean square difference 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 heart rate.

9. The system according to claim 1, characterized in that, The aforementioned feature vector construction based on the frequency domain of EDA signals involves: using the Welch method for power spectral density estimation, and subdividing the Tonic and Phasic components into five frequency bands within a recommended frequency range; the Welch method divides the time series into multiple overlapping windows, and then applies a Fourier transform to each window to obtain the PSD estimate for each window; the obtained PSD of each window 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; based on this, for each predefined frequency band, the statistical characteristics of its spectral power are calculated, including mean, standard deviation, median, kurtosis, skewness, maximum and minimum values; 10. The HRV signal is decomposed into three frequency bands based on different frequency ranges: high frequency (0.15 Hz to 0.4 Hz), low frequency (0.04 Hz to 0.15 Hz), and extremely low frequency (0.003 Hz to 0.04 Hz). The Lomb-Scargle method is used to estimate the power spectral density, and the following characteristic parameters are extracted: extremely low frequency power, low frequency power, high frequency power, total power, and the LF / HF ratio. The formula for the Lomb-Scargle periodogram method is as follows: ;in, , It is each frequency to be tested. It is the number of data points. It is the timestamp of each data point. At any moment The measured value, yes The average value, yes variance This is the time offset.

Images