An early judgment system for Parkinson's disease cognitive impairment based on multi-dimensional features of transcranial magnetic stimulation - electroencephalogram

Abstract

The present invention discloses an early judgment system for Parkinson's disease cognitive impairment based on multi-dimensional features of transcranial magnetic stimulation - electroencephalogram, including: a transcranial magnetic stimulation - electroencephalogram combined acquisition module, a synchronous signal preprocessing module, a multi-dimensional feature extraction module, a feature intelligent fusion module, and a storage and display module; the transcranial magnetic stimulation - electroencephalogram combined acquisition module is used to encode and perturb the patient's nerve activity through single-pulse transcranial magnetic stimulation and collect the evoked electroencephalogram signals; the synchronous signal preprocessing module is used to denoise and filter the evoked electroencephalogram signals; the multi-dimensional feature extraction module uses multi-dimensional electroencephalogram feature analysis to extract features from the processed evoked electroencephalogram signals to obtain multi-parameter electroencephalogram nerve activity features; the feature intelligent fusion module is used to assign weights and fuse the multi-parameter electroencephalogram nerve activity features to obtain multi-modal nerve indexes for cognitive impairment; the storage and display module is used to store the multi-modal nerve indexes for cognitive impairment and visualize the multi-modal nerve indexes for cognitive impairment.

Description

Technical Field

[0001] This invention belongs to the fields of brain-computer interface technology and medical signal technology, specifically involving an early diagnosis system for cognitive impairment in Parkinson's disease based on multidimensional features of transcranial magnetic-electroencephalography (TME). Background Technology

[0002] Cognitive impairment in Parkinson's disease is characterized by a progressive decline in cognitive function, which worsens with age, disease duration, and severity. Clinically, it is heterogeneous, involving deficits in one or more cognitive domains, including executive function impairment, accompanied by deficits in language, visuospatial abilities, and memory. Cognitive impairment is prevalent in both the early and late stages of Parkinson's disease and carries a high risk of developing Parkinson's dementia. Its insidious onset significantly impacts quality of life and increases the social and healthcare burden. In recent years, the incidence of cognitive impairment in Parkinson's disease has been on the rise, becoming an increasingly serious public health problem.

[0003] Currently, there is no effective treatment for cognitive impairment in Parkinson's disease, making early assessment crucial for slowing disease progression. Existing resting-state EEG characteristics, such as time-frequency oscillatory activity and alterations in brain network functional connectivity, are associated with cognitive decline in Parkinson's patients. However, resting-state EEG indicators are easily interfered with by patients' speech or motor symptoms, and imaging-based cortical deficit indicators are not yet apparent in the early stages of cognitive decline. Previous early assessments of cognitive impairment in Parkinson's disease have been based on single-dimensional neurological characteristics, lacking stable multi-parameter neurological indicators. Transcranial magnetic stimulation combined with electroencephalography (TMS-EEG), as an emerging brain activity assessment method, can non-invasively and directly record cortical activation induced by TMS in specific cortical regions, reflecting multi-dimensional cortical responses such as cortical excitability, oscillatory dynamics, and effective connectivity. It requires no subject participation, offering advantages such as operational safety, stable and reproducible indicators. Combined with multi-dimensional EEG characteristic analysis methods, it can serve as an effective means for early assessment of cognitive impairment in Parkinson's disease.

[0004] Therefore, this study specifically designed an early assessment system for cognitive impairment in Parkinson's disease patients based on transcranial magnetic stimulation-electroencephalography (TMS-EEG). Summary of the Invention

[0005] This invention aims to address the shortcomings of existing technologies and provides the following solutions:

[0006] An early diagnostic system for cognitive impairment in Parkinson's disease based on multidimensional features of transcranial magnetic resonance imaging (TME) and electroencephalography (EEG) includes: a combined TME acquisition module, a synchronous signal preprocessing module, a multidimensional feature extraction module, a feature intelligent fusion module, and a storage and display module;

[0007] The transcranial magnetic-electroencephalogram (TME) combined acquisition module is used to perturb the patient's neural activity through single-pulse transcranial magnetic stimulation and to acquire evoked EEG signals.

[0008] The synchronization signal preprocessing module is used to perform noise reduction filtering on the evoked EEG signal to obtain the processed evoked EEG signal.

[0009] The multidimensional feature extraction module uses multidimensional EEG feature analysis to extract features from the processed evoked EEG signal to obtain multi-parameter EEG neural activity features, which include: time-domain features, time-frequency domain features, and multi-dimensional brain network neural features.

[0010] The feature intelligent fusion module is used to construct a fusion feature machine learning model, and to use the fusion feature machine learning model to perform weight allocation and feature fusion on the multi-parameter EEG neural activity features to obtain multimodal neural indicators of cognitive impairment.

[0011] The storage and display module is used to store the cognitive impairment multimodal neural indicators and visualize them.

[0012] Preferably, the transcranial magnetic stimulation-electroencephalogram (TMS-EEG) combined acquisition module includes: a resting motion threshold measurement unit, a transcranial magnetic stimulation unit, and an EEG acquisition unit;

[0013] The resting motion threshold measurement unit is used to apply transcranial magnetic stimulation to the primary motor cortex on the side of the patient's disease at a 45° anterior-posterior angle, causing muscle contraction on the contralateral side to generate motor evoked potentials, and the minimum intensity required to cause at least 5 electromyographic peak amplitudes greater than or equal to 50 μV in 10 consecutive stimulations is used as the resting motion threshold.

[0014] The transcranial magnetic stimulation unit is used to apply 80 single pulses tangentially to the right posterior parietal cortex of the patient at a 45° anterior-posterior angle, with each single pulse stimulation interval being 2 to 4 seconds, and the pulse intensity being set to 90% of the resting exercise threshold.

[0015] The EEG acquisition unit acquires the evoked EEG signals at a sampling rate of 1000Hz.

[0016] Preferably, the workflow of the synchronization signal preprocessing module includes:

[0017] Import the evoked EEG signals and transcranial magnetic stimulation electrode information, locate the electrodes based on the electrode information, and delete useless electrodes;

[0018] The events of interest in the evoked EEG signals are extracted and the corresponding evoked EEG signals are locked. Then, baseline correction is performed to remove transcranial magnetic stimulation pulse artifacts and evoked electromyography artifacts.

[0019] A first missing data interpolation was performed around the transcranial magnetic stimulation pulse to minimize artifacts caused by low-pass and anti-aliasing filters. Then, the interpolated data of the first missing data interpolation was replaced with first constant amplitude EEG signal data. Then, independent component analysis and automatic component selection were used to remove voltage attenuation artifacts caused by scalp muscle activation and transcranial magnetic stimulation induced electromyography in the evoked EEG signal.

[0020] The evoked EEG signal within 2ms before and 10ms after the transcranial magnetic stimulation pulse is deleted. A second missing data interpolation is performed around the transcranial magnetic stimulation pulse, followed by bandpass filtering of 1-100Hz and bandstop filtering of 48-52Hz. Then, the interpolated data of the second missing data interpolation is replaced with the second constant EEG signal amplitude data. Finally, independent component analysis and automatic component selection are used again to remove the non-transcranial magnetic stimulation lock artifacts remaining in the evoked EEG signal.

[0021] A third missing data interpolation is performed around the transcranial magnetic stimulation pulse. The electrode channel CPz of the missing EEG signal data is interpolated, and the electrode information is rereferenced to a common average reference to obtain the processed evoked EEG signal.

[0022] Preferably, the multidimensional feature extraction module includes: a time-domain feature extraction unit, a time-frequency domain feature extraction unit, and a brain network feature extraction unit;

[0023] The time-domain feature extraction unit is used to extract the global average field power and transcranial magnetic stimulation evoked potentials of each electrode at each time point in the processed evoked EEG signal.

[0024] The time-frequency domain feature extraction unit is used to extract the average power of each frequency band of each electrode channel of the whole brain in the processed evoked EEG signal within a certain time period.

[0025] The brain network feature extraction unit is used to extract the weighted phase lag index matrix and brain network attribute graph theory parameters of the processed evoked EEG signal.

[0026] Preferably, the workflow of the temporal feature extraction unit includes:

[0027] The global average field power was calculated by evaluating the standard deviation of the amplitude of the EEG signal at each electrode at each time point in the evoked EEG signal after the treatment.

[0028]

[0029] Where n represents the total number of EEG signal channels, i represents the number of EEG signal channels, and A i (t) represents the amplitude of the EEG signal in the i-th channel, A mean (t) represents the average amplitude of all channels at time t;

[0030] The peak values ​​of transcranial magnetic stimulation (TMS) evoked potentials (TMPs) of each EEG signal channel within five time windows of interest in the processed evoked EEG signal were extracted as TMS evoked potentials, namely P30, N45, P60, N100, and P180.

[0031]

[0032] Where x represents each subject, P_tep represents the average peak value of the event in all channels of all subjects during the time period from t1 to t2. P_tep represents positive potentials P30, P60 and P180, and N_tep represents negative potentials N45 and N100.

[0033] Preferably, the workflow of the time-frequency domain feature extraction unit includes:

[0034] Molay wavelet convolution was used to perform time-frequency decomposition on the five frequency bands δ, θ, α, β and γ in the processed evoked EEG signal to extract the signal amplitude and obtain the energy at specific time and frequency points. Then, the time window was divided into 100ms intervals, and the average power of each frequency band of each electrode channel in the whole brain in the processed evoked EEG signal during the period from 0 to 400ms after transcranial magnetic stimulation was calculated.

[0035] Preferably, the workflow of the brain network feature extraction unit includes:

[0036] Based on the time-frequency matrix after time-frequency decomposition, the phase information of each electrode in the processed evoked EEG signal is extracted, and the phase difference matrix is ​​obtained by subtracting the phase matrices. The cross spectrum between electrodes is obtained, and the weighted phase lag index matrix is ​​calculated based on the cross spectrum.

[0037]

[0038] Where Im[·] denotes taking the imaginary part, G i，j (t, f) represents the cross spectrum of two time series of the i-th EEG signal channel and the j-th EEG signal channel, where t represents time and f represents frequency;

[0039] Graph theory methods are used to quantitatively describe the characteristics of brain networks, and graph theory parameters of the brain network attributes are calculated. These parameters include: clustering coefficient, average shortest path length, global efficiency, and local efficiency.

[0040]

[0041] Where CC represents the clustering coefficient, ASPL represents the average shortest path length, GE represents global efficiency, LE represents local efficiency, N represents the number of nodes in the brain network, and h a f is the number of nodes adjacent to node a. a It is h a The actual number of connecting edges between nodes, l ab G is the shortest path length between node a and node b. a It is a subgraph consisting of the adjacent nodes of node a, l bh It is the shortest path length between node b and node h.

[0042] Preferably, the feature intelligent fusion module includes: a data processing unit, a feature selection unit, a weight analysis unit, a model building unit, and a weighted fusion unit;

[0043] The data processing unit is used to preprocess the time-domain features, the time-frequency domain features, and the multi-dimensional brain neural features of the brain network to obtain the processed features.

[0044] The feature selection unit uses an elastic network regression feature selection method to determine the optimal feature selection parameters;

[0045] The weight analysis unit is used to assign weights to the time-domain features, the time-frequency features, and the multi-dimensional brain neural features of the brain network in the optimal feature selection parameters, so as to obtain the weight ratio of different features.

[0046] The model building unit constructs a multi-kernel support vector machine fusion model by combining different kernel functions to obtain the fusion feature machine learning model;

[0047] The weighted fusion unit uses the fusion feature machine learning model to match features of different kernel functions with features of different dimensions, and fuses the features of different kernel functions based on the weight ratio to obtain the cognitive impairment multimodal neural index.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. In response to the problem of unstable resting-state EEG indicators in existing inventions, this invention independently designed a single-pulse transcranial magnetic stimulation coding scheme, selected the right posterior parietal cortex, a brain region integrating cognition and motor function, as the magnetic stimulation target, and simultaneously collected evoked EEG signals of subjects under disturbed conditions, which improved the objectivity of disease cognitive function detection and could more accurately reflect the symptom characteristics of mild cognitive impairment in Parkinson's disease.

[0050] 2. This invention utilizes a time-frequency-space multidimensional brain nerve feature extraction algorithm, which, compared with existing single-dimensional EEG features, can provide multi-parameter EEG features of cortical excitability, neural oscillation dynamics, and brain network connectivity, thereby increasing the feature dimensions and reliability of indicators for Parkinson's disease with mild cognitive impairment.

[0051] 3. This invention uses artificial intelligence algorithms to construct a fusion feature machine learning model, which differs from existing identification methods based on quantitative electroencephalography. It can optimize disease feature selection, determine the weight allocation of multimodal disease features, remove redundant information, and improve the identification specificity of mild cognitive impairment in Parkinson's disease patients. Attached Figure Description

[0052] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of the transcranial magnetic-electroencephalogram (TME) combined acquisition module according to an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of the experimental results of an embodiment of the present invention. In this diagram, A represents the average butterfly diagram of the transcranial magnetic stimulation evoked potential waveforms of all electrode channels, B represents the time-frequency diagram of the average whole-brain oscillation power evoked after transcranial magnetic stimulation, and C represents the statistical difference in brain network functional connectivity between two groups constructed by weighted phase lag indices in the Theta and Alpha frequency bands. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] Example 1

[0059] In this embodiment, as Figure 1 As shown, an early diagnosis system for cognitive impairment in Parkinson's disease based on multidimensional features of transcranial magnetic resonance imaging (TME) includes: a combined TME acquisition module, a synchronous signal preprocessing module, a multidimensional feature extraction module, a feature intelligent fusion module, and a storage and display module.

[0060] The transcranial magnetic stimulation-electroencephalogram (TMS-EEG) combined acquisition module is used to perturb the patient's neural activity through single-pulse transcranial magnetic stimulation and to acquire evoked EEG signals.

[0061] The transcranial magnetic stimulation (TMS)-electroencephalography (EEG) combined acquisition module includes: a resting motor threshold measurement unit, a TMS unit, and an EEG acquisition unit. The resting motor threshold measurement unit is used to apply TMS to the primary motor cortex on the affected side of the patient at a 45° anteroposterior angle, inducing contralateral muscle contraction to generate motor evoked potentials. The minimum intensity required to induce at least 5 peak-to-peak amplitudes of electromyography (EMG) greater than or equal to 50 μV in 10 consecutive stimulations is defined as the resting motor threshold. The TMS unit is used to apply 80 single pulses to the right posterior parietal cortex of the patient at a 45° anteroposterior angle, with each single pulse stimulation interval of 2–4 seconds and the pulse intensity set to 90% of the resting motor threshold. The EEG acquisition unit acquires evoked EEG signals at a sampling rate of 1000 Hz.

[0062] In this embodiment, as Figure 2 As shown, the transcranial magnetic stimulation (TMS)-electroencephalography (EEG) combined acquisition module includes a TMS device and a synchronous EEG acquisition device. The TMS device includes a frameless computerized stereotactic neuronavigation system, an electromyography (EMG) acquisition device, a magnetic stimulation generator, and a rapid biphasic magnetic stimulator. The EEG acquisition device includes an EEG cap and an EEG signal acquisition device. The frameless computerized stereotactic neuronavigation system is responsible for tracking and recording the coil position. The EMG acquisition device is responsible for online monitoring of EMG activity to determine the resting motor threshold. The magnetic stimulation generator is responsible for generating magnetic stimulation with selected parameters. TMS is performed by a rapid biphasic magnetic stimulator equipped with a figure-8 coil (70mm) to transmit TMS pulses. The synchronous EEG acquisition device uses a 64-channel EEG with clinically compliant wet electrodes to continuously record evoked EEG signals. Ag / AgCl electrodes are used, and electrode placement follows the 10-20 international standard. This device is used to synchronously acquire EEG signals evoked by TMS in real time, and a display screen synchronously displays the acquired EEG signals.

[0063] In this embodiment, the workflow of the transcranial magnetic stimulation-electroencephalogram (TMS-EEG) combined acquisition module includes: (1) navigation and positioning: the position of the coil is tracked and recorded based on the frameless computerized stereotactic neuronavigation system, and the signal acquisition brain region and electrode position are determined by the combined TMS-compatible 64-channel EEG system; (2) acquisition instruction: Parkinson's disease patients are asked to sit in a chair with their forearms on the armrests, remain relaxed, and look at the "black cross" on the computer screen 70 cm in front of them; (3) resting motor threshold measurement: TMS is applied to the primary motor cortex on the side of the patient's disease at a 45° anterior-posterior angle, causing the contralateral muscles to contract and generate motor evoked potentials, which are obtained by connecting the Ag / AgCl surface electrodes to the abdominal tendon, wherein the active electrode is placed on the muscle belly and the reference electrode is placed on the metacarpophalangeal joint of the index finger. Electromyography activity was continuously monitored online, and the resting motion threshold was defined as the minimum intensity required to induce an interpeak amplitude of ≥50μV in at least 5 out of 10 consecutive tests; (4) Transcranial magnetic stimulation synchronous EEG acquisition: 80 single pulses were applied to the right posterior parietal cortex of the patient at a 45° anterior-posterior angle, with an interval of 2-4 seconds between each single pulse, and the pulse intensity was set to 90% of the resting motion threshold; (5) Whole brain EEG signals were acquired at a sampling rate of 1000Hz, and the impedance value was kept below 5kΩ.

[0064] The synchronization signal preprocessing module is used to denoise and filter the evoked EEG signal to obtain the processed evoked EEG signal.

[0065] The workflow of the synchronized signal preprocessing module includes: importing evoked EEG signals and transcranial magnetic stimulation (TMS) electrode information; locating electrodes based on the electrode information and deleting useless electrodes; extracting events of interest from the evoked EEG signals and locking the corresponding evoked EEG signals; then performing baseline correction to remove TMS pulse artifacts and evoked EMG artifacts; performing a first missing data interpolation around the TMS pulse to minimize artifacts caused by low-pass and anti-aliasing filters; then replacing the interpolated data of the first missing data interpolation with first constant amplitude EEG signal data; then using independent component analysis and automatic component selection to remove voltage attenuation artifacts caused by scalp muscle activation and TMS-evoked EMG in the evoked EEG signals; and finally processing the data from the first 2ms to the last 2ms of the TMS pulse. The evoked EEG signal within 10ms is deleted to minimize artifacts caused by low-pass and anti-aliasing filters. A second missing data interpolation is performed around the transcranial magnetic stimulation pulse, followed by bandpass filtering of 1–100Hz and bandstop filtering of 48–52Hz. Then, the interpolated data of the second missing data interpolation is replaced with second constant amplitude data. Then, independent component analysis and automatic component selection are used again to remove residual non-transcranial magnetic stimulation locking artifacts in the evoked EEG signal. A third missing data interpolation is performed around the transcranial magnetic stimulation pulse. The electrode channel CPz of the missing EEG signal data is interpolated, that is, the average value of the EEG signal of the surrounding electrodes is used as its EEG signal, and the electrode information is rereferenced as a common average reference to obtain the processed evoked EEG signal.

[0066] The multidimensional feature extraction module uses multidimensional EEG feature analysis to extract features from the processed evoked EEG signals, obtaining multi-parameter EEG neural activity features, which include: time-domain features, time-frequency domain features, and multi-dimensional brain network neural features.

[0067] The multidimensional feature extraction module includes: a time-domain feature extraction unit, a time-frequency domain feature extraction unit, and a brain network feature extraction unit.

[0068] The temporal feature extraction unit is used to extract the global average field power and transcranial magnetic stimulation evoked potentials of each electrode at each time point in the processed evoked EEG signal.

[0069] The workflow of the temporal feature extraction unit includes: calculating the global average field power by evaluating the standard deviation of the amplitude of the EEG signal at each electrode at each time point in the evoked EEG signal after processing.

[0070]

[0071] Where n represents the total number of EEG signal channels, i represents the number of EEG signal channels, and A i (t) represents the amplitude of the EEG signal in the i-th channel, A mean(t) represents the average amplitude of all channels at time t; the peak values ​​of transcranial magnetic stimulation evoked potentials (TMS) of each EEG signal channel within the five time windows of interest in the processed evoked EEG signal are extracted as TMS, namely P30 (15-35ms), N45 (35-55ms), P60 (55-75ms), N100 (75-150ms) and P180 (150-250ms):

[0072]

[0073] Where x represents each subject, P_tep represents the average peak value of the event in all channels of all subjects during the time period from t1 to t2. P_tep represents positive potentials P30, P60 and P180, and N_tep represents negative potentials N45 and N100.

[0074] The time-frequency domain feature extraction unit is used to extract the average power of each frequency band of each electrode channel in the whole brain of the processed evoked EEG signal within a certain time period.

[0075] The workflow of the time-frequency domain feature extraction unit includes: using Molay wavelet convolution to perform time-frequency decomposition on five frequency bands in the processed evoked EEG signal: δ (1-4Hz), θ (4-7Hz), α (8-13Hz), β (14-30Hz), and γ (31-45Hz), extracting the signal amplitude, and obtaining the energy at specific time-frequency points. The parameters of the Molay wavelet convolution include: bandwidth parameter (Fb) of 7, center frequency parameter of 0.5, and total scale of 1024. The calculation formula is as follows:

[0076]

[0077] Where Sca represents the scale value, Fc represents the center frequency parameter, Tsc represents the total scale parameter, m represents the scale parameter, w represents the intermediate variable, and ψ(t) represents the Molay wavelet function; then the time window is divided into 100ms intervals, and the average power of each frequency band of each electrode channel in the evoked EEG signal after processing from 0 to 400ms after transcranial magnetic stimulation is calculated.

[0078] The brain network feature extraction unit is used to extract the weighted phase lag index matrix and brain network attribute graph theory parameters of the processed evoked EEG signals.

[0079] The workflow of the brain network feature extraction unit includes: calculating the weighted phase lag index to quantify the signal coupling degree between channels; establishing a brain network with scalp electrodes as network nodes and the weighted phase lag index as connection parameters; extracting the phase information of each electrode in the evoked state EEG signal after time-frequency decomposition based on the time-frequency matrix, subtracting the phase matrices to obtain the phase difference matrix, obtaining the cross spectrum between electrodes, and calculating the weighted phase lag index matrix based on the cross spectrum.

[0080]

[0081] Where Im[·] denotes taking the imaginary part, G i，j (t, f) represents the cross spectrum of two time series of the i-th and j-th EEG signals, where t represents time and f represents frequency. Graph theory is used to quantitatively describe the characteristics of the brain network, calculating graph theory parameters of brain network attributes. These parameters include: clustering coefficient, average shortest path length (ASPL), global efficiency, and local efficiency. In this embodiment, the clustering coefficient (CC) is constructed to measure the probability that adjacent nodes of a given node in the network are adjacent to each other. The average shortest path length (ASPL), which is the reciprocal of each other, and the global efficiency (GE) jointly reflect the efficiency of information transmission within the network. Simultaneously, local efficiency (LE) is constructed to quantify the ability to transmit brain interaction information in local parts of the brain network.

[0082]

[0083] Where CC represents the clustering coefficient, ASPL represents the average shortest path length, GE represents global efficiency, LE represents local efficiency, N represents the number of nodes in the brain network, and h a f is the number of nodes adjacent to node a. a It is h a The actual number of connecting edges between nodes, l ab G is the shortest path length between node a and node b. a It is a subgraph consisting of the adjacent nodes of node a, l bh It is the shortest path length between node b and node h.

[0084] The feature intelligent fusion module is used to construct a fusion feature machine learning model, and to use the fusion feature machine learning model to perform weight allocation and feature fusion on multi-parameter EEG neural activity features to obtain multimodal neural indices of cognitive impairment.

[0085] The feature intelligent fusion module includes a data processing unit, a feature selection unit, a weight analysis unit, a model building unit, and a weighted fusion unit. The data processing unit preprocesses temporal features, time-frequency features, and multi-dimensional neuronal features of the brain network to obtain processed features. The feature selection unit uses an elastic network regression feature selection method to determine the optimal feature selection parameters. The weight analysis unit assigns weights to temporal features, time-frequency features, and multi-dimensional neuronal features of the brain network in the optimal feature selection parameters to obtain the weight proportions of different features. The model building unit constructs a multi-kernel support vector machine fusion model by combining different kernel functions to obtain a fused feature machine learning model. The weighted fusion unit uses the fused feature machine learning model to match different kernel functions with features of different dimensions and fuses the features of different kernel functions based on their weight proportions to obtain multimodal neurological indicators of cognitive impairment.

[0086] In this embodiment, the data processing unit performs feature matrix construction, outlier removal, data normalization, and sign testing on time-domain features, time-frequency features, and multi-dimensional brain neural features of the brain network, respectively. Specifically, feature matrices are constructed based on various features such as global average field strength, transcranial magnetic stimulation evoked potentials, weighted phase lag exponent, and graph theory parameters, and training label vectors are set for each type of feature matrix. Outlier values ​​in the feature matrix, such as null values ​​and infinitely large values, are removed to avoid computer memory overflow leading to model training failure. Zero-mean normalization is used for data normalization to avoid the influence of dimensionality issues between different feature data on feature weight calculation. Sign testing is used for preliminary feature screening to avoid individual errors being introduced into the model. The feature selection unit adopts the elastic network regression feature selection method, setting two layers of parameter optimization iteration to determine the optimal feature selection parameters. Specifically, one layer of regularization is used for feature selection and model simplification, and the second layer of regularization prevents model overfitting. During the iteration process, the weight coefficients corresponding to each feature are trained, while features with zero coefficients, i.e., redundant and invalid features, are removed to improve the iteration speed. The weighting analysis unit introduces weight parameters during the construction of the multi-kernel support vector machine model to measure the weight proportion of spatiotemporal multidimensional features in identifying cognitive impairment in Parkinson's disease, i.e., the contribution (weight) of different dimensional features to the classification result of the fusion model. To obtain weight parameters more accurately, cross-validation is used in conjunction with a parameter optimization strategy. The optimal feature selection parameters are determined through internal iteration, and the optimal parameter classification model and its weight parameters are extracted. To measure the importance of sub-features under each dimension, randomized repeated 8-fold cross-validation is used. The optimal feature selection result corresponding to the optimal fusion classifier obtained in each test group is statistically analyzed, and the frequency of each dimension feature is calculated and sorted from high to low to determine the feature weight of each feature vector. The model construction unit constructs a multi-kernel support vector machine fusion model by combining different kernel functions to obtain a fusion feature machine learning model. The weighted fusion unit uses the fusion feature machine learning model to perform feature matching between different kernel functions and features of different dimensions, and fuses the features of different kernel functions based on the weight proportion to obtain multimodal neural indices of cognitive impairment.

[0087] The storage and display module is used to store multimodal neural indicators of cognitive impairment and to visualize these indicators.

[0088] Example 2

[0089] In this embodiment, to verify the validity and effectiveness of this application, the following scientific experiments were conducted, specifically including:

[0090] (1) Subject Enrollment: 45 Parkinson's disease patients (21 males, 24 females, 62.11 ± 6.95 years old) were invited to participate in the experiment. All subjects signed informed consent forms before enrollment. The inclusion criteria were as follows: 1) Hoehn-Yahr stage ≤ 3; 2) Stable vital signs, no obvious cardiopulmonary or osteoarthritis diseases; 3) Stable medication, no adjustment in the past 3 months; 4) No other special treatment required during hospitalization; 5) No deep brain stimulation or implanted therapy; 6) Able to understand the informed consent form, willing to sign it, and committed to completing the assessment and treatment. The exclusion criteria were: 1) Patients with fractures or mental symptoms; 2) Patients with clinical dementia score > 0.5 or visual or hearing impairments; 3) Severe resting tremor; 4) Presence of metal implants, such as pacemakers and brain pacemakers; 5) History of epilepsy; 6) Other serious systemic diseases.

[0091] All enrolled participants were categorized according to the diagnostic criteria for Level I mild cognitive impairment in Parkinson's disease developed by the Movement Disorders Association Working Group. Parkinson's disease patients with a Montreal Cognitive Assessment (MoCA) score ≤25 were assigned to the mild cognitive impairment group (n=22), while those with a MoCA score >25 were assigned to the cognitively normal group (n=23).

[0092] (2) Experimental process: Using the judgment system of the present invention, transcranial magnetic-electroencephalography (TME) was sequentially used to complete transcranial magnetic-electroencephalography (TME) acquisition, synchronous signal preprocessing, multi-feature extraction, and intelligent fusion of disease features for all Parkinson's disease patients.

[0093] (3) Experimental results: such as Figure 3 As shown, Figure 3 (A) The average butterfly diagram of transcranial magnetic stimulation evoked potential waveforms from all electrode channels in the PD-NC group (left) and PD-MCI group (right). The EEG neural activity feature components used to construct the fusion feature machine learning model, such as P30 and N45, have been marked in the figure. Figure 3 (B) Time-frequency plots of the mean power of whole-brain oscillations induced by transcranial magnetic stimulation in the PD-NC group (left) and PD-MCI group (right). The dashed box highlights the EEG neural activity feature components used to construct the fusion feature machine learning model, such as the alpha band of 0-100ms and theta band of 200-300ms. Figure 3 (C) shows the statistical differences in functional connectivity of brain networks constructed using weighted phase lag indices in the Theta and Alpha bands. Bars indicate electrode connections with significant differences, serving as EEG neural activity feature components that will be used to construct a fusion-feature machine learning model.

[0094] like Figure 3As shown, in terms of cortical activity, PD-NC and PD-MCI patients exhibited significant differences in the P30, P60, P180, N45, and N100 components of the time-domain TEP, with PD-MCI patients showing stronger cortical activity. Time-frequency results revealed significant differences between PD-NC and PD-MCI patients primarily in the theta (4-7Hz) and alpha (8-13Hz) oscillations. Brain network analysis showed significant differences in network connectivity between the frontal, parietal, and central lobes in PD-NC and PD-MCI patients. These results can provide effective potential multidimensional neurological indicators for the early identification of cognitive impairment in clinical Parkinson's disease.

[0095] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An early diagnostic system for cognitive impairment in Parkinson's disease based on multidimensional transcranial magnetic-electroencephalography (TME), characterized in that, include: The transcranial magnetic-electroencephalogram (TME) acquisition module, synchronous signal preprocessing module, multidimensional feature extraction module, feature intelligent fusion module, and storage and display module are included. The transcranial magnetic-electroencephalogram (TME) combined acquisition module is used to perturb the patient's neural activity through single-pulse transcranial magnetic stimulation and to acquire evoked EEG signals. The synchronization signal preprocessing module is used to perform noise reduction filtering on the evoked EEG signal to obtain the processed evoked EEG signal. The multidimensional feature extraction module uses multidimensional EEG feature analysis to extract features from the processed evoked EEG signal to obtain multi-parameter EEG neural activity features, which include: time-domain features, time-frequency domain features, and multi-dimensional brain network neural features. The feature intelligent fusion module is used to construct a fusion feature machine learning model, and to use the fusion feature machine learning model to perform weight allocation and feature fusion on the multi-parameter EEG neural activity features to obtain multimodal neural indicators of cognitive impairment. The storage and display module is used to store the cognitive impairment multimodal neural indicators and visualize the cognitive impairment multimodal neural indicators; The transcranial magnetic-electroencephalogram (TME) acquisition module includes: a resting motion threshold measurement unit, a transcranial magnetic stimulation unit, and an electroencephalogram (EEG) acquisition unit; The resting motion threshold measurement unit is used to apply transcranial magnetic stimulation to the primary motor cortex on the side of the patient's disease at a 45° anterior-posterior angle, causing muscle contraction on the contralateral side to generate motor evoked potentials, and the minimum intensity required to cause at least 5 electromyographic peak amplitudes greater than or equal to 50 μV in 10 consecutive stimulations is used as the resting motion threshold. The transcranial magnetic stimulation unit is used to apply 80 single pulses tangentially to the patient's right posterior parietal cortex at a 45° anterior-posterior angle, with each single pulse interval being 2-4 seconds, and the pulse intensity being set to 90% of the resting motor threshold. The EEG acquisition unit acquires the evoked EEG signals at a sampling rate of 1000Hz; The workflow of the synchronization signal preprocessing module includes: Import the evoked EEG signals and transcranial magnetic stimulation electrode information, locate the electrodes based on the electrode information, and delete useless electrodes; The events of interest in the evoked EEG signals are extracted and the corresponding evoked EEG signals are locked. Then, baseline correction is performed to remove transcranial magnetic stimulation pulse artifacts and evoked electromyography artifacts. A first missing data interpolation was performed around the transcranial magnetic stimulation pulse to minimize artifacts caused by low-pass and anti-aliasing filters. Then, the interpolated data of the first missing data interpolation was replaced with first constant amplitude EEG signal data. Then, independent component analysis and automatic component selection were used to remove voltage attenuation artifacts caused by scalp muscle activation and transcranial magnetic stimulation induced electromyography in the evoked EEG signal. The evoked EEG signal within 2 ms before and 10 ms after the transcranial magnetic stimulation pulse is deleted. A second missing data interpolation is performed around the transcranial magnetic stimulation pulse, followed by bandpass filtering of 1–100 Hz and bandstop filtering of 48–52 Hz. Then, the interpolated data of the second missing data interpolation is replaced with the second constant EEG signal amplitude data. Finally, independent component analysis and automatic component selection are used again to remove the non-transcranial magnetic stimulation lock artifacts remaining in the evoked EEG signal. A third missing data interpolation is performed around the transcranial magnetic stimulation pulse. The electrode channel CPz of the missing EEG signal data is interpolated, and the electrode information is rereferenced to a common average reference to obtain the processed evoked EEG signal. The multidimensional feature extraction module includes: a time-domain feature extraction unit, a time-frequency domain feature extraction unit, and a brain network feature extraction unit; The time-domain feature extraction unit is used to extract the global average field power and transcranial magnetic stimulation evoked potentials of each electrode at each time point in the processed evoked EEG signal. The time-frequency domain feature extraction unit is used to extract the average power of each frequency band of each electrode channel of the whole brain in the processed evoked EEG signal within a certain time period. The brain network feature extraction unit is used to extract the weighted phase lag exponent matrix and brain network attribute graph theory parameters of the processed evoked EEG signal; The workflow of the temporal feature extraction unit includes: The global average field power was calculated by evaluating the standard deviation of the amplitude of the EEG signal at each electrode at each time point in the evoked EEG signal after the treatment. Where n represents the total number of EEG signal channels, i represents the number of EEG signal channels, and A i (t) represents the amplitude of the EEG signal in the i-th channel, A mean (t) represents the average amplitude of all channels at time t; The peak values ​​of transcranial magnetic stimulation (TMS) evoked potentials (TMPs) of each EEG signal channel within five time windows of interest in the processed evoked EEG signal were extracted as TMS evoked potentials, namely P30, N45, P60, N100, and P180. Where x represents each subject, tep represents the average peak value of event potentials in all channels of all subjects during the time period from t1 to t2. P_tep represents positive potentials P30, P60 and P180, and N_tep represents negative potentials N45 and N100. The workflow of the time-frequency domain feature extraction unit includes: Molay wavelet convolution was used to perform time-frequency decomposition on the five frequency bands δ, θ, α, β and γ in the processed evoked EEG signal to extract the signal amplitude and obtain the energy at specific time and frequency points. Then, the time window was divided into 100ms intervals, and the average power of each frequency band of each electrode channel in the whole brain in the processed evoked EEG signal during the period from 0 to 400ms after transcranial magnetic stimulation was calculated. The workflow of the brain network feature extraction unit includes: Based on the time-frequency matrix after time-frequency decomposition, the phase information of each electrode in the processed evoked EEG signal is extracted, and the phase difference matrix is ​​obtained by subtracting the phase matrices. The cross spectrum between electrodes is obtained, and the weighted phase lag index matrix is ​​calculated based on the cross spectrum. Where Im[·] denotes taking the imaginary part, The cross spectrum represents the two time series of the i-th EEG signal channel and the j-th EEG signal channel, where t represents time and f represents frequency; Graph theory methods are used to quantitatively describe the characteristics of brain networks, and graph theory parameters of the brain network attributes are calculated. These parameters include: clustering coefficient, average shortest path length, global efficiency, and local efficiency. Where CC represents the clustering coefficient, ASPL represents the average shortest path length, GE represents global efficiency, LE represents local efficiency, N represents the number of nodes in the brain network, and h a f is the number of nodes adjacent to node a. a It is h a The actual number of connecting edges between nodes, l ab G is the shortest path length between node a and node b. a It is a subgraph consisting of the adjacent nodes of node a, l bh It is the shortest path length between node b and node h; The feature intelligent fusion module includes: a data processing unit, a feature selection unit, a weight analysis unit, a model building unit, and a weighted fusion unit; The data processing unit is used to preprocess the time-domain features, the time-frequency domain features, and the multi-dimensional brain neural features of the brain network to obtain the processed features. The feature selection unit uses an elastic network regression feature selection method to determine the optimal feature selection parameters; The weight analysis unit is used to assign weights to the time-domain features, time-frequency features, and multi-dimensional brain neural features of the brain network in the optimal feature selection parameters, so as to obtain the weight ratio of different features. The model building unit constructs a multi-kernel support vector machine fusion model by combining different kernel functions to obtain the fusion feature machine learning model; The weighted fusion unit uses the fusion feature machine learning model to match features of different kernel functions with features of different dimensions, and fuses the features of different kernel functions based on the weight ratio to obtain the cognitive impairment multimodal neural index.

Images