Method, apparatus and processor for realizing methamphetamine addiction identification and craving determination on basis of machine learning, and computer-readable storage medium therefor

Abstract

The present invention relates to a method for realizing methamphetamine addiction identification and craving determination on the basis of machine learning. The method comprises the following steps: collecting and pre-processing high-density scalp resting-state electroencephalogram data of a subject to obtain an EEG time series; on the basis of a standard head model, performing source locating on the acquired time series data; using a pre-defined brain template to perform electroencephalogram functional connectivity analysis; using a functional connectivity network to detect the temporal correlation or phase synchronization between a plurality of brain regions; and performing real craving manifestation feature determination on the subject by means of establishing and verifying a machine learning model. The present invention also relates to a corresponding apparatus, a processor and a storage medium therefor. By means of the method, apparatus and processor for realizing methamphetamine addiction identification and craving determination on the basis of machine learning and the storage medium therefor in the present invention, accurate identification of a methamphetamine addict and a craving state thereof can be expected, thereby facilitating the assessment of a treatment effect on methamphetamine addicts, and reducing the recurrence rate.

Description

Method, device, processor and computer-readable storage medium for realizing methamphetamine addiction identification and craving discrimination based on machine learning

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to the Chinese invention patent application No. 202311529756.0 filed on November 16, 2023, the contents of which are incorporated by reference into this application. Technical Field

[0003] The present invention relates to the field of biomedical technology, in particular to the field of neuroscience technology, and specifically to a method, device, processor, and computer-readable storage medium thereof for realizing methamphetamine addiction identification and craving discrimination based on machine learning. Background Art

[0004] [Corrected 16.10.2024 in accordance with Regulation 26] Meth addicts exhibit persistent psychological dependence (high craving) and behavioral traits such as impulsivity, which contribute to high-risk behaviors such as relapse and violence, placing a heavy burden on family economics and social stability. Craving is the direct cause of relapse. Identifying EEG markers of craving in methamphetamine addicts can serve as objective markers of relapse during the withdrawal period, crucially assisting physicians in the diagnosis, monitoring, treatment, and prognosis of methamphetamine addicts. Scalp electroencephalography (EEG), with its high temporal resolution, low cost, and ease of data acquisition, offers a viable approach for identifying biomarkers of craving in methamphetamine addicts. However, conventional 32- or 64-channel scalp EEG performs poorly in source localization analysis, where volume conduction can significantly alter the scalp EEG signal. Notably, the use of high-density EEG (128 channels or higher) can significantly mitigate the effects of volume conduction. High-density EEG systems provide source reconstruction of both regional and global brain activity, through which brain activity can be reconstructed by applying precise source imaging.

[0005] Existing techniques: 1. Hypothesis-driven, where researchers compare different groups (e.g., a group of methamphetamine addicts and a group of healthy controls) to explore the clinical relevance of specific neural signals based on a prior hypothesis. For example, previous studies have shown that increased slow wave activity (delta and theta) and increased beta band synchronization in the medial prefrontal cortex (MPFC) are associated with increased cued craving in methamphetamine abusers. In this approach, prior EEG evidence is used to determine the neural indicators to be measured, the population to be tested, and the analytical methods to be performed. Using this approach, although multiple brain regions have been identified as associated with methamphetamine craving and can distinguish between healthy controls and methamphetamine addicts, these associations and statistically significant differences have not resulted in strong biomarkers for methamphetamine addicts. Hypothesis-driven approaches attempt to elucidate the variance or differences between multiple variables, but have limited applicability to new individuals (1,2);

[0006] 2. Data-driven machine learning methods that use various neurobiological features to classify methamphetamine addicts and predict craving. Previously, the mean and standard deviation of absolute EEG power and galvanic skin response were proposed to classify methamphetamine addicts and healthy controls (3,4), but the relatively low-density EEG poses challenges in effectively modeling the topological characteristics of input connections. To date, high-density EEG biomarkers have not been successfully used to predict methamphetamine addiction and reliably identify methamphetamine addicts at the individual level.

[0007] Summary of the Invention

[0008] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method, device, processor and computer-readable storage medium thereof for realizing methamphetamine addiction identification and craving discrimination based on machine learning.

[0009] To achieve the above objectives, the present invention provides a method, device, processor, and computer-readable storage medium for identifying methamphetamine addiction and craving based on machine learning.

[0010] The main features of the method for identifying methamphetamine addiction and craving based on machine learning are as follows:

[0011] (1) Collect high-density scalp resting-state EEG data from the subjects;

[0012] (2) Preprocessing the collected data to remove noise data and artifacts to obtain the preprocessed EEG time series;

[0013] (3) Based on the standard head model, the acquired time series data are processed using a predefined brain template to perform brain region source localization processing;

[0014] (4) Using functional connectivity networks to detect temporal correlations or phase synchronization between multiple brain regions;

[0015] (5) By establishing and validating a machine learning model, the characteristics of the subjects’ true craving expression are identified.

[0016] Preferably, the step (1) is specifically as follows:

[0017] All 128 scalp channels of the high-density EEG cap were adjusted to maintain an impedance of <50 kΩ and a sampling rate of 500 Hz. The collected resting-state EEG data generated by methamphetamine-induced cues were filtered offline using a 0.01-100 Hz bandpass filter or a 50 Hz notch filter.

[0018] Preferably, the step (2) specifically includes the following steps:

[0019] (2.1) Offline preprocessing of the collected high-density EEG data was performed by downsampling the time series to 250 Hz, removing the 50 Hz power frequency noise through a notch filter, and bandpass filtering between 1 Hz and 100 Hz through a zero-phase finite impulse response filter;

[0020] (2.2) Remove bad channels including bridged channels and noise channels and perform spherical interpolation. Paroxysmal segments are excluded and removed. Independent component analysis is used to decompose independent components to remove artifacts such as blinking, eye movement, heartbeat, and electromyography.

[0021] (2.3) Use principal component analysis to detect and remove undesirable components through dimensionality reduction, and determine the minimum number of principal components that accounts for more than 99.9% of the total variance as the dimension to obtain the preprocessed EEG time series;

[0022] (2.4) Data preprocessing was completed by filtering the sensor-level preprocessed time series into five typical frequency bands: delta (1–3 Hz), theta (4–7 Hz), alpha (8–12 Hz), beta (13–30 Hz), and gamma (31–50 Hz) through a custom pipeline and verification by EEG experts.

[0023] Preferably, the step (3) specifically includes the following steps:

[0024] (3.1) Using OpenMEEG to calculate a three-layer symmetric boundary element model of the head, we selected 3003 rotated dipoles from the cortical surface and estimated the imaging kernel using a depth-weighted and regularized minimum norm estimation method. The kernel was mapped from the scalp EEG to cortical source-level current density. Principal component analysis was then used to reduce the current density time series of each vertex in three spatial dimensions from three orthogonal axes to one dimension.

[0025] (3.2) Generate a time series representing the main change patterns of all vertices in the same region of interest through principal component analysis, extract the dominant signal of the time series using singular value decomposition, and regard the first singular vector as the signal at the region of interest level;

[0026] The 31 regions of interest described in (3.3) were independently divided according to functional connectivity of resting-state fMRI, including the VIS (visual network): bilateral primary visual area (V1); somatosensory network: bilateral primary somatosensory cortex (S1); DAN (dorsal attention network): inferior frontal junction (IFJ), intraparietal sulcus (IPS), frontal visual field (FEF) and accessory visual field (SEF); DMN (default mode network): posterior cingulate cortex (PCC), medial prefrontal cortex (MPFC) and angular gyrus (ANG); FPN (frontoparietal control network): posterior middle frontal gyrus (PMFG), inferior parietal lobe (IPL), orbital gyrus (ORB) and middle temporal gyrus (MTG); and VAN (ventral attention network): prefrontal middle gyrus (AMFG), insular gyrus (INS), dorsal anterior cingulate cortex (DACC) and supramarginal gyrus (SUP).

[0027] Preferably, the step (4) is specifically as follows:

[0028] The imaginary coherence connectivity of 465 unique ROI pairs was calculated for each of the five frequency bands and the two resting eye conditions to obtain the interactions between the brain regions of interest at the cortical level. ij,t (f) is expressed as x _i (t) and x _j (t) is the cross-spectral density at time point or trial t, where S ij,t (f) is expressed as:

[0029] S ij,t (f)= <x i,t (f)x j,t * (f)>,

[0030] Where * represents the complex conjugate, and <> represents the expected value;

[0031] S _i,j(f) is expressed as a cross spectrum, and the imaginary part consistency is defined as:

[0032] Preferably, the step (5) specifically includes the following steps:

[0033] (5.1) Perform group-level analysis on the source-space FCN and use the false discovery rate adjustment to correct the p-value of each pair of functional connections;

[0034] (5.2) Use a biomarker-based machine learning model for craving prediction in addicted individuals and identification of abnormal addicted individuals to assess the strength of the relationship between the two biomarkers;

[0035] (5.3) Use an association vector machine with a high-density linear kernel for EEG data to build a sparse linear regression model for predicting craving scores from connectivity features, and use an RVM with a high-density linear kernel for EEG data and a sigmoid function outer layer to build a sparse classification model to distinguish methamphetamine addicts from healthy subjects;

[0036] (5.4) Use the RVM machine learning model to maximize the marginal likelihood with sparse priors to penalize complex models, thereby automatically selecting relevant features for prediction;

[0037] (5.5) Craving prediction for methamphetamine-addicted individuals: An RVM model with 10-fold 10-fold cross validation was used to predict craving scores in methamphetamine-addicted individuals based on brain functional connectivity in the brain regions of interest;

[0038] (5.6) Abnormal identification of methamphetamine-addicted individuals: Based on brain density EEG functional connectivity estimation, an RVM model with 10-fold 10-fold cross-validation was used to classify methamphetamine-addicted individuals and healthy control individuals. The model assigned the label "1" to individuals with methamphetamine addiction and the label "0" to individuals in the healthy control group. The classification performance was evaluated using classification accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve;

[0039] (5.7) Abnormal identification and craving prediction using sensor spatial power: The power spectrum of each frequency band in a single brain region of interest is used as the input feature of the machine learning model, and the classification and prediction performance between cortical functional connectivity and frequency band power spectrum are compared.

[0040] The main features of the device for identifying methamphetamine addiction and craving based on machine learning are as follows:

[0041] a processor configured to execute computer-executable instructions;

[0042] A memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the method for identifying methamphetamine addiction and determining craving based on machine learning.

[0043] The main feature of the processor for implementing methamphetamine addiction identification and craving discrimination based on machine learning is that the processor is configured to execute computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for implementing methamphetamine addiction identification and craving discrimination based on machine learning are implemented.

[0044] The main feature of the computer-readable storage medium is that a computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of the above-mentioned method for identifying methamphetamine addiction and determining craving based on machine learning.

[0045] The method, device, processor, and computer-readable storage medium thereof for realizing methamphetamine addiction identification and craving discrimination based on machine learning of the present invention have brought significant innovations to the diagnosis and treatment of methamphetamine addiction by combining high-density scalp EEG data with analysis of neural connection networks, and provided a new approach for treatment methods based on the individual level, which is expected to improve the treatment effect of methamphetamine addicts and reduce the relapse rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] FIG1 is a schematic diagram of high-density EEG data acquisition and feature extraction according to the method for realizing methamphetamine addiction identification and craving discrimination based on machine learning of the present invention.

[0047] FIG2 is a schematic diagram of establishing and verifying a machine learning model for the method for realizing methamphetamine addiction identification and craving discrimination based on machine learning of the present invention. DETAILED DESCRIPTION

[0048] In order to more clearly describe the technical content of the present invention, further description is given below in conjunction with specific embodiments.

[0049] Before describing in detail embodiments according to the present invention, it should be noted that, hereinafter, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, whereby a process, method, article, or apparatus comprising a list of elements includes not only those elements, but also other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0050] This technical solution specifically consists of the following solutions:

[0051] High-density EEG data acquisition and feature extraction:

[0052] (1) Recruiting methamphetamine addicts and healthy subjects;

[0053] (2) High-density scalp EEG data acquisition, preprocessing, and source localization analysis;

[0054] (3) Obtain the resting state brain activity of 31 regions and EEG functional connectivity between 31 brain regions;

[0055] Building and validating machine learning models into software:

[0056] (4) Using a correlation vector machine to establish an EEG feature machine learning model to predict methamphetamine craving;

[0057] (5) Logistic regression was used to establish an EEG feature machine learning model to identify methamphetamine addicts;

[0058] (6) Both models were validated on additional datasets.

[0059] Referring to FIG1 , the method for identifying methamphetamine addiction and craving based on machine learning includes the following steps:

[0060] (1) Collect high-density scalp resting-state EEG data from the subjects;

[0061] (2) Preprocessing the collected data to remove noise data and artifacts to obtain the preprocessed EEG time series;

[0062] (3) Based on the standard head model, the acquired time series data are processed using a predefined brain template to perform brain region source localization processing;

[0063] (4) Using functional connectivity networks to detect temporal correlations or phase synchronization between multiple brain regions;

[0064] (5) By establishing and validating a machine learning model, the characteristics of the subjects’ true craving expression are identified.

[0065] As a preferred embodiment of the present invention, the step (1) is specifically as follows:

[0066] All 128 scalp channels of the high-density EEG cap were adjusted to maintain an impedance of <50 kΩ and a sampling rate of 500 Hz. The collected resting-state EEG data generated by methamphetamine-induced cues were filtered offline using a 0.01-100 Hz bandpass filter or a 50 Hz notch filter.

[0067] As a preferred embodiment of the present invention, the step (2) specifically includes the following steps:

[0068] (2.1) Offline preprocessing of the collected high-density EEG data was performed by downsampling the time series to 250 Hz, removing the 50 Hz power frequency noise through a notch filter, and bandpass filtering between 1 Hz and 100 Hz through a zero-phase finite impulse response filter;

[0069] (2.2) Bad channels, including bridged and noisy channels, were removed and spherical interpolation was performed. Paroxysmal segments were excluded and excised. Independent component analysis was used to decompose the independent components, which include activity originating from muscle motor unit groups (e.g., in the eyes) and strong, high-frequency, broadband activity that clusters many action potentials during muscle contraction and static tension. Specifically, the interpolated EEG signal was subjected to modal decomposition. The time series recorded on the scalp is a spatially stable mixture of temporally independent brain and artifact activity. A "decomposition" matrix W was first found that decomposes or linearly decomposes the multi-channel scalp data into a sum of temporally independent and spatially stationary components. The rows of the output data matrix, U = WX, are the time courses of activation of the principal component components. The columns of the inverse matrix, inv(W), give the relative projected intensities of each component at each scalp sensor (right). These scalp weights assign scalp topography to each component and provide evidence for their physiological origins.

[0070] (2.3) artifact components were detected using principal component analysis, and the scalp topography of the four selected components and the artifact-corrected EEG time series obtained by removing the four selected electrooculogram and muscle noise components from the data were analyzed;

[0071] (2.4) Data preprocessing was completed by filtering the sensor-level preprocessed time series into five typical frequency bands: delta (1–3 Hz), theta (4–7 Hz), alpha (8–12 Hz), beta (13–30 Hz), and gamma (31–50 Hz) through a custom pipeline and verification by EEG experts.

[0072] As a preferred embodiment of the present invention, the step (3) specifically includes the following steps:

[0073] (3.1) Using OpenMEEG to calculate a three-layer symmetric boundary element model of the head, we selected 3003 rotated dipoles from the cortical surface and estimated the imaging kernel using a depth-weighted and regularized minimum norm estimation method. The kernel was mapped from the scalp EEG to cortical source-level current density. Principal component analysis was then used to reduce the current density time series of each vertex in three spatial dimensions from three orthogonal axes to one dimension.

[0074] (3.2) Generate a time series representing the main change patterns of all vertices in the same region of interest through principal component analysis, extract the dominant signal of the time series using singular value decomposition, and regard the first singular vector as the signal at the region of interest level;

[0075] The 31 regions of interest described in (3.3) were independently divided according to functional connectivity of resting-state fMRI, including the VIS (visual network): bilateral primary visual area (V1); somatosensory network: bilateral primary somatosensory cortex (S1); DAN (dorsal attention network): inferior frontal junction (IFJ), intraparietal sulcus (IPS), frontal visual field (FEF) and accessory visual field (SEF); DMN (default mode network): posterior cingulate cortex (PCC), medial prefrontal cortex (MPFC) and angular gyrus (ANG); FPN (frontoparietal control network): posterior middle frontal gyrus (PMFG), inferior parietal lobe (IPL), orbital gyrus (ORB) and middle temporal gyrus (MTG); and VAN (ventral attention network): prefrontal middle gyrus (AMFG), insular gyrus (INS), dorsal anterior cingulate cortex (DACC) and supramarginal gyrus (SUP).

[0076] As a preferred embodiment of the present invention, the step (4) is specifically as follows:

[0077] The imaginary coherence was used to calculate the connectivity of 465 (31x31) pairs of brain regions of interest in each of the five frequency bands and the two resting eye conditions to obtain the interactions between the brain regions of interest at the cortical level. ij,t (f) is expressed as x _i (t) and x _j (t) is the cross-spectral density at time point or trial t, where S ij,t (f) is expressed as:

[0078] S ij,t (f)= <x i,t (f)x j,t * (f)>,

[0079] Where * represents the complex conjugate, <> represents the expected value, and x _i (t) is the signal of brain area i at a certain time point, x _j (t) is the signal of brain area j at time point.

[0080] S _i,j (f) is expressed as a cross spectrum, and the imaginary part consistency is defined as:

[0081] Calculation method: The imaginary part of the electrode cross-spectrum is divided by the square root of the power of each brain area, where S ij(f) is the cross spectrum of brain area i and brain area j, S ii (f) and S jj are the powers of each brain region.

[0082] As a preferred embodiment of the present invention, the step (6) specifically includes the following steps:

[0083] (5.1) Group-level analysis of the imaginary consistency of source-space functional connectivity was performed, and the p-value of each pair of functional connections was corrected using the false discovery rate adjustment;

[0084] (5.2) Use a biomarker-based machine learning model for craving prediction in addicted individuals and identification of abnormal addicted individuals to assess the strength of the relationship between the two biomarkers;

[0085] (5.3) Use an association vector machine with a high-density linear kernel for EEG data to build a sparse linear regression model for predicting craving scores from connectivity features, and use an RVM with a high-density linear kernel for EEG data and a sigmoid function outer layer to build a sparse classification model to distinguish methamphetamine addicts from healthy subjects;

[0086] (5.4) Use the RVM machine learning model to maximize the marginal likelihood with sparse priors to penalize complex models, thereby automatically selecting relevant features for prediction;

[0087] (5.5) Craving prediction in methamphetamine-addicted individuals: An RVM model with 10-fold 10-fold cross-validation was used to predict craving scores in methamphetamine-addicted individuals based on brain functional connectivity in the brain regions of interest;

[0088] (5.6) Abnormal identification of methamphetamine-addicted individuals: Based on brain density EEG functional connectivity estimation, an RVM model with 10-fold 10-fold cross-validation was used to classify methamphetamine-addicted individuals and healthy control individuals. The model assigned the label "1" to individuals with methamphetamine addiction and the label "0" to healthy control individuals. Classification performance was evaluated using classification accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve.

[0089] (5.7) Abnormality identification and craving prediction using source-level power: The power spectra of each frequency band in a single brain region of interest are used as input features of the machine learning model, and the classification and prediction performance between electrode functional connectivity and frequency band power spectra are compared.

[0090] In practical applications, this technical solution aims to use high-density scalp EEG data for craving prediction and anomaly identification. It analyzes high-density scalp EEG data by calculating cortical functional connectivity after source localization to achieve the following two main goals: (1) predict the level of methamphetamine craving in methamphetamine addicts, and (2) identify differences between methamphetamine addicts and healthy individuals.

[0091] Working steps and technical details of implementation:

[0092] 1. Data acquisition: All 128 scalp channels were adjusted to maintain an impedance of <50 kΩ. The sampling rate was 500 Hz, and offline filtering was performed using a 0.01-100 Hz bandpass filter (using a 50 Hz notch filter). Participants were asked to record resting-state high-density EEG after watching a video of methamphetamine-inducing cues (5 minutes): sitting in a quiet room with eyes closed for 5 minutes, remaining alert and relaxed, then staring at a given point with eyes open for 5 minutes, and then closing their eyes for another 5 minutes. Custom MATLAB scripts were used for preprocessing and source localization.

[0093] 2. Preprocessing: Offline preprocessing of high-density EEG data was performed. The time series was downsampled to 250 Hz, 50 Hz power-frequency noise was removed using notch filtering, and bandpass filtering was performed between 1 Hz and 100 Hz using a zero-phase finite impulse response filter. Bad channels, including bridged and noisy channels, were then removed and spherical interpolation was performed. Paroxysmal segments were excluded and excised. To remove artifacts such as blinks, eye movements, heartbeats, and myoelectricity, independent component analysis (ICA) with an extended information maximization algorithm was used to decompose the independent components. Principal component analysis (PCA) was then used to detect and remove bad components through dimensionality reduction, and the minimum number of principal components accounting for more than 99.9% of the total variance was determined as the dimensionality. After the final re-referencing, the preprocessed EEG time series was obtained. Through a custom pipeline and validation by EEG experts, an automated algorithm was used to detect and remove extreme values ​​of flat amplitudes or outliers, channel correlation, and robust paroxysmal spike measurements. The preprocessed time series at the sensor level are filtered into five typical frequency bands: delta (1–3 Hz), theta (4–7 Hz), alpha (8–12 Hz), beta (13–30 Hz), and gamma (31–50 Hz).

[0094] 3. Source localization: Source localization was performed based on a standard head model. A three-layer symmetric boundary element model of the head was calculated using OpenMEEG, and then a rotated dipole of 3003 vertices on the cortical surface was selected. The minimum norm estimation method (MNE) with depth weighting and regularization was used to estimate the imaging kernel, which was mapped from the scalp electrode EEG to the cortical source-level current source density. The current density time series of each vertex in three spatial dimensions was then reduced from three orthogonal axes to one dimension using principal component analysis.

[0095] Connectivity analysis was performed in 31 regions of interest (ROIs) of the Montreal Neurological Institute (MNI)-defined brain template, derived from independent parcellation of functional connectivity from resting-state fMRI clustered by independent component analysis (ICA). Principal component analysis was used to generate time series representing the dominant pattern of change across all vertices within the same ROI. Singular value decomposition (SVD) was used to extract the dominant signal of the time series, with the first singular vector considered as the signal at the ROI level. These 31 regions of interest were divided into six brain functional networks: VIS (visual network): bilateral primary visual area (V1); somatosensory network: bilateral primary somatosensory cortex (S1); DAN (dorsal attention network): inferior frontal junction (IFJ), intraparietal sulcus (IPS), frontal visual field (FEF) and accessory visual field area (SEF); DMN (default mode network): posterior cingulate cortex (PCC), medial prefrontal cortex (MPFC) and angular gyrus (ANG); FPN (frontoparietal control network): posterior middle frontal gyrus (PMFG), inferior parietal lobe (IPL), orbital gyrus (ORB) and middle temporal gyrus (MTG); and VAN (ventral attention network): prefrontal middle gyrus (AMFG), insular gyrus (INS), dorsal anterior cingulate cortex (DACC) and supramarginal gyrus (SUP).

[0096] 5. Functional connectivity network calculation: Functional connectivity networks (FCNs) play a vital role in capturing dynamic interactions by detecting temporal correlation or phase synchronization between multiple brain regions. Coherence and coherence are two widely used methods for analyzing EEG connectivity. In order to reduce the influence of volume conduction of scalp signals, a solution involving the extraction of imaginary coherence was developed, which represents the time delay between two channels. In this study, imaginary coherence (iCoh) was used to reveal cortical interactions. Here, iCoh connectivity was calculated for 465 unique ROI pairs in each of the 5 frequency bands and in each of the 2 resting eye conditions.

[0097] In this technical solution, we use imaginary coherence (iCoh) to reveal the interactions between brain regions of interest at the cortical level. The calculation method is: the imaginary part of the electrode cross-spectrum divided by the square root of the power of each brain region

[0098] Denote S(ij,t) as the cross spectral density of x_i(t) and x_j(t) at time point or trial t. S(ij,t) is defined as

[0099] S ij,t (f)= <x i,t (f)x j,t * (f)>,

[0100] Where * represents the complex conjugate and <> represents the expected value.

[0101] Expressing S_ij(f) as a cross spectrum, the imaginary consistency is defined as

[0102] Due to the antisymmetry of the imaginary part, calculating this connectivity at the region of interest level by averaging the correlation of each source space vertex between two given regions of interest nodes may result in an error [iCoh]_ij(f)=-[iCoh]_ji(f)). Therefore, to avoid this error, iCoh at the region of interest level in the carrier frequency band was directly calculated.

[0103] 6. Group analysis:

[0104] (1) Group-level analyses of source-space FCN (functional connectivity, imaginary coherence iCoh) were performed. To investigate the relationship between craving and FCN, we calculated the correlation between these two factors, while correcting the p-values ​​for multiple comparisons using false discovery rate (FDR) adjustment. In addition, we performed an independent t-test to compare the FCN between the methamphetamine-addicted group and the healthy control group, and corrected the p-values ​​for each pair of functional connections using FDR adjustment. In addition, we calculated the correlation between craving and behavioral measures, while correcting them using FDR adjustment.

[0105] (2) Machine Learning Models: Two machine learning models based on two biomarkers (prediction of craving in addicted individuals and identification of abnormal addicted individuals) were used to assess the strength of this relationship. For both models, functional connectivity measures between each brain region of interest were used. See the figure for an overview of the machine learning workflow.

[0106] (3) Relevance Vector Machine: A relevance vector machine (RVM) with a high-density linear kernel for EEG data was used to build a sparse linear regression model for predicting craving scores from connectivity features. An RVM with a high-density linear kernel for EEG data and a sigmoid function outer layer was also used to build a sparse classification model to distinguish methamphetamine addicts from healthy subjects.

[0107] (4) RVM is a machine learning technique that automatically selects relevant features for prediction by penalizing complex models by maximizing the marginal likelihood with sparse priors in a sparse Bayesian learning framework. Compared with support vector machines (SVMs), RVMs can provide probabilistic predictions, reduce computational complexity, and estimate error / margin trade-off parameters. In addition, the kernel function of RVM does not need to satisfy the Mercer condition, which means that it can determine hyperparameters without additional validation.

[0108] (5) Craving prediction in methamphetamine-addicted individuals: An RVM model with 10 10-fold cross-validation (the data were randomly divided into 10 subsets, each containing approximately the same number of subjects. One subset was selected as the test data, and the other 9 subsets were used as the training data. This process was repeated 10 times, and each of the 10 subsets was used once as the test data. Therefore, each subject had a predicted probability. To improve the stability of the prediction, the data were randomly divided into 10 groups, and stratified 10-fold cross-validation was performed on each randomized dataset) was used to predict craving scores in methamphetamine-addicted individuals based on brain functional connectivity in the brain regions of interest. We evaluated the prediction performance by calculating the Pearson correlation coefficient and root mean square error (RMSE) between the cross-validated predicted scores and the true scores. We also report the p-value of the one-tailed alternative hypothesis that the coefficient of determination of iCoh in the five bands was greater than 0 (a total of 10 patterns) under resting eyes-open and eyes-closed conditions after correction for multiple comparisons using the FDR method. In addition, since the RVM model is a sparse model, it has the ability to use the feature weight matrix of the RVM model to indicate the predictive importance of brain functional connectivity.

[0109] (6) Abnormal identification of methamphetamine-addicted individuals: Based on brain density EEG functional connectivity estimation, an RVM model with 10 10-fold cross validation was used to classify methamphetamine-addicted individuals and healthy control individuals. The model assigned the label "1" to individuals with methamphetamine addiction and the label "0" to individuals in the healthy control group. The classification performance was evaluated using classification accuracy, sensitivity (true positive rate), specificity (true negative rate), and area under the receiver operating characteristic curve (AUC). Higher AUC values ​​indicate better separability and higher classification performance. The diagnostic odds ratio, defined as the ratio of the odds of testing positive when the subject has the disease to the odds of testing positive when the subject does not have the disease, was also calculated to measure the effectiveness of the classification model.

[0110] (7) Abnormal identification and craving prediction using sensor spatial power: To test the predictive ability of brain functional connectivity in methamphetamine-addicted individuals, the power spectrum of each frequency band in the spatial space of a single brain region of interest was used as the input feature of the machine learning model, and the classification and prediction performance between cortical functional connectivity and frequency band power spectrum were compared.

[0111] In practical application, this technical solution has the following significant technical effects:

[0112] 1. EEG feature machine learning models predict methamphetamine craving; models based on REOβ and REOδ can significantly predict methamphetamine craving scores. Specifically, the REOδ model has a Pearson's r of 0.41, p = 0.009 (FDR correction), and a root mean square error of 21.60, while the REOβ model shows the highest performance, with a Pearson's r of 0.73 and p < 3×10 -9 (FDR corrected), and the RMS error = 14.75.

[0113] 2. EEG feature machine learning model for identifying methamphetamine addicts: evaluated through 10 10-fold cross-validation. The REOβ model showed the best classification performance across all frequency bands and resting conditions, achieving a classification accuracy of 80.95% (sensitivity: 82.14%; specificity: 79.59%), significantly higher than other models (all models had classification accuracy below 70%). In addition, compared with other frequency bands and resting conditions, the REOβ model also showed the highest AUC (0.859) and diagnostic odds ratio (DOR, 17.94).

[0114] We tested whether the craving prediction model could be replicated by performing the same analysis on an independent sample of methamphetamine addicts’ induced resting-state EEG dataset. First, we tested the model performance of the best predictor, REO beta, trained on dataset 1 using dataset 2, and observed a relatively high level of performance (Pearson’s r = 0.60, p < 2 × 10 -5 , RMSE = 19.31), which is only slightly lower than the performance on the training dataset. This result demonstrates the robustness of the craving prediction model trained on REOβ of the previous dataset. Next, the methamphetamine addiction dataset 2 was used as an independent dataset and 10 10-fold cross-validations were performed. The predictive performance of the model established for all frequency bands and resting conditions showed that the REOβ model was able to significantly predict the craving score (Pearson's r = 0.65, p < 2 × 10 -5, FDR corrected, RMSE = 17.56), which is consistent with the results of the methamphetamine addict dataset. However, the REO-delta model did not show significant prediction compared with the previous dataset (Pearson's r = 0.29, p = 0.222, FDR corrected, RMSE = 22.78). Consistent with the results obtained with the discovery dataset, the results of the REO-beta model trained on the methamphetamine addict dataset 2 demonstrated the robustness of the trained RVM model in predicting methamphetamine addict craving. Finally, the weights of the RVM model of REOβ trained on dataset 2 were evaluated. As with the weights trained on dataset 1, it was observed that the connections between the medial prefrontal cortex and the right superior limbic gyrus (e.g., prefrontal cortex-right superior limbic gyrus) played an important role in the prediction.

[0115] Individual-level biomarkers: The invention of a machine learning algorithm provides a biomarker that can measure craving at the individual level, which is important in the diagnosis and treatment of methamphetamine addiction because the degree of craving can vary from patient to patient.

[0116] Application of scalp EEG data: This technical solution innovatively utilizes high-density scalp EEG data and combines it with the craving-related brain functional connection network to achieve more accurate diagnosis and monitoring.

[0117] Clinical pathways for personalized treatment: This technology provides a new approach to treatment at the individual level, which is expected to improve treatment outcomes and reduce relapse rates for methamphetamine addicts.

[0118] The device for identifying methamphetamine addiction and craving based on machine learning comprises:

[0119] a processor configured to execute computer-executable instructions;

[0120] A memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the method for identifying methamphetamine addiction and determining craving based on machine learning.

[0121] The processor for implementing methamphetamine addiction identification and craving discrimination based on machine learning, wherein the processor is configured to execute computer-executable instructions, and when the computer-executable instructions are executed by the processor, the steps of the method for implementing methamphetamine addiction identification and craving discrimination based on machine learning are implemented.

[0122] The computer-readable storage medium stores a computer program thereon, and the computer program can be executed by a processor to implement the steps of the method for identifying methamphetamine addiction and determining craving based on machine learning.

[0123] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0124] It should be understood that each part of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution device.

[0125] Those skilled in the art will understand that all or part of the steps of the method for implementing the above-mentioned embodiment can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0126] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0127] Throughout this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "embodiment" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0128] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

[0129] The method, device, processor, and computer-readable storage medium thereof for realizing methamphetamine addiction identification and craving discrimination based on machine learning of the present invention have brought significant innovations to the diagnosis and treatment of methamphetamine addiction by combining high-density scalp EEG data with analysis of neural connection networks, and provided a new approach for treatment methods based on the individual level, which is expected to improve the treatment effect of methamphetamine addicts and reduce the relapse rate.

[0130] In this specification, the present invention has been described with reference to specific embodiments thereof. However, it will be apparent that various modifications and variations may be made without departing from the spirit and scope of the present invention. Accordingly, the specification and drawings are to be regarded as illustrative rather than restrictive.

Claims

1. A method for identifying methamphetamine addiction and craving based on machine learning, characterized in that: The method comprises the following steps: (1) Collect high-density scalp resting-state EEG data from the subjects; (2) preprocessing the collected data to remove noise data and artifacts to obtain the preprocessed EEG time series; (3) Based on the standard head model, the acquired time series data is processed using a predefined brain template to perform brain region source localization processing; (4) Use functional connectivity networks to detect temporal correlations or phase synchronization between multiple brain regions; (5) By establishing and verifying a machine learning model, the characteristics of the subjects' true craving behavior are identified.

2. The method for realizing methamphetamine addiction identification and craving discrimination based on machine learning according to claim 1, characterized in that: The step (1) is specifically as follows: All 128 scalp channels of the high-density EEG cap were adjusted to keep their impedance at <50 kΩ, with a sampling rate of 500 Hz. The collected resting-state EEG data generated by methamphetamine-induced cues were filtered offline using a 0.01-100 Hz bandpass filter or a 50 Hz notch filter.

3. The method for realizing methamphetamine addiction identification and craving discrimination based on machine learning according to claim 1, characterized in that: The step (2) specifically comprises the following steps: (2.1) The collected high-density EEG data were preprocessed offline by downsampling the time series to 250 Hz, removing the 50 Hz power frequency noise through a notch filter, and performing bandpass filtering between 1 Hz and 100 Hz through a zero-phase finite impulse response filter; (2.2) Remove bad channels including bridged channels and noisy channels and perform spherical interpolation. Paroxysmal segments are excluded and removed. Independent component analysis is used to decompose independent components to remove artifacts such as blinking, eye movement, heartbeat, and electromyography. (2.3) Use principal component analysis to detect and remove bad components through dimensionality reduction, and determine the minimum number of principal components that accounts for more than 99.9% of the total variance as the dimension to obtain the preprocessed EEG time series; (2.4) Data preprocessing was completed by filtering the sensor-level preprocessed time series into five typical frequency bands: delta (1-3 Hz), theta (4-7 Hz), alpha (8-12 Hz), beta (13-30 Hz), and gamma (31-50 Hz) through a custom pipeline and verification by EEG experts.

4. The method for realizing methamphetamine addiction identification and craving discrimination based on machine learning according to claim 3, characterized in that: The step (3) specifically comprises the following steps: (3.1) Use OpenMEEG to calculate the three-layer symmetric boundary element model of the head, then select the rotated dipoles of 3003 vertices on the cortical surface, and use the minimum norm estimation method with depth weighting and regularization to estimate the imaging kernel, and the imaging kernel is mapped from the scalp electrode EEG to the cortical source-level current source density. Then, the current density time series of each vertex in three spatial dimensions is reduced from three orthogonal axes to one dimension through principal component analysis; (3.2) Generate a time series representing the main change patterns of all vertices in the same region of interest through principal component analysis, extract the dominant signal of the time series using singular value decomposition, and regard the first singular vector as the signal at the region of interest level; The 31 regions of interest described in (3.3) were independently divided according to the functional connectivity of resting-state fMRI, including VIS (visual network): bilateral primary visual area (V1); somatosensory network: bilateral primary somatosensory cortex (S1); DAN (dorsal attention network): inferior frontal junction (IFJ), intraparietal sulcus (IPS), frontal visual field (FEF) and supplementary visual field (SEF); DMN (default mode network): posterior cingulate cortex (PCC), medial prefrontal cortex (MPFC) and angular gyrus (ANG); FPN (frontoparietal control network): posterior middle frontal gyrus (PMFG), inferior parietal lobe (IPL), orbital gyrus (ORB) and middle temporal gyrus (MTG); and VAN (ventral attention network): prefrontal middle gyrus (AMFG), insular gyrus (INS), dorsal anterior cingulate cortex (DACC) and supramarginal gyrus (SUP).

5. The method for realizing methamphetamine addiction identification and craving discrimination based on machine learning according to claim 4, characterized in that: The step (4) is specifically as follows: The connectivity of the imaginary coherence was calculated for 465 unique ROI pairs in each of the five frequency bands and the two resting eye conditions to obtain interactions between the brain regions of interest at the cortical level. ij,t (f) is represented by x _i (t) and x _j (t) is the cross-spectral density at time point or trial t, where S ij,t (f) is expressed as: S ij,t (f)=<x i,t (f)x j,t * (f)>, Among them, * represents complex conjugate, <> represents expected value; S _i,j (f) is expressed as a cross spectrum, and the imaginary consistency is defined as: Among them, S ij (f) is the cross spectrum of brain area i and brain area j, S ii (f) and S jj are the powers of each brain region.

6. The method for realizing methamphetamine addiction identification and craving discrimination based on machine learning according to claim 5, characterized in that: The step (5) specifically comprises the following steps: (5.1) Perform group-level analysis on source-space FCN and use false discovery rate adjustment to correct the p-value of each pair of functional connections; (5.2) Use a biomarker-based machine learning model for craving prediction in addicted individuals and identification of abnormal addicted individuals to assess the strength of the relationship between the two biomarkers; (5.3) Use the association vector machine with a high-density EEG data linear kernel to build a sparse linear regression model to obtain The concatenated features predicted the craving score and used RVM with a high-density EEG data linear kernel and a sigmoid function outer layer to build a sparse classification model to distinguish between methamphetamine addicts and healthy subjects; (5.4) Use the RVM machine learning model to maximize the marginal likelihood with sparse priors to penalize complex models, thereby automatically selecting relevant features for prediction; (5.5) Craving prediction for methamphetamine addicted individuals: The RVM model with 10 10-fold cross validation was used to predict craving scores in methamphetamine addicted individuals based on brain functional connectivity in the brain regions of interest; (5.6) Abnormal identification of methamphetamine-addicted individuals: Based on brain density EEG functional connectivity estimation, the methamphetamine-addicted individuals and healthy control group individuals were classified using the RVM model with 10 10-fold cross validation. The model assigned the label "1" to individuals with methamphetamine addiction and the label "0" to individuals in the healthy control group. The classification performance was evaluated using classification accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve; (5.7) Abnormality identification and craving prediction using sensor spatial power: The power spectrum of each frequency band in a single brain region of interest is used as the input feature of the machine learning model, and the classification and prediction performance between cortical functional connectivity and frequency band power spectrum are compared.

7. A device for identifying methamphetamine addiction and craving based on machine learning, characterized in that: The device comprises: a processor configured to execute computer executable instructions; A memory storing one or more computer executable instructions, wherein when the computer executable instructions are executed by the processor, the steps of the method for realizing methamphetamine addiction identification and craving discrimination based on machine learning as described in any one of claims 1 to 6 are implemented.

8. A processor for realizing methamphetamine addiction identification and craving discrimination based on machine learning, characterized in that: The processor is configured to execute computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the method for realizing methamphetamine addiction identification and craving discrimination based on machine learning as described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of the method for realizing methamphetamine addiction identification and craving discrimination based on machine learning as described in any one of claims 1 to 6.

Images