Non-suicidal self-injury identification and prediction method and system

Abstract

The present invention relates to a method for identifying and predicting non-suicidal self-injury, comprising: collecting behavioral data and EEG signals from a subject; preprocessing the collected EEG signals; analyzing the preprocessed EEG signals; and constructing a classification model based on the collected behavioral data and analyzed EEG signals to identify and predict non-suicidal self-injury. The present invention also relates to a system for identifying and predicting non-suicidal self-injury. This invention can provide an objective and accurate detection method, enabling early identification and monitoring of NSSI, and creating favorable conditions for timely intervention and treatment.

Description

Technical Field

[0001] The present invention relates to a method and system for identifying and predicting non-suicidal self-injury. Background Art

[0002] Nonsuicidal self-injury (NSSI) refers to recurring behaviors in which individuals intentionally harm their own tissues and organs without suicidal intent. It is prevalent among adolescents aged 10 to 19, with a prevalence rate as high as 20% to 30%. It poses a serious threat to adolescents' physical and mental health. Effective identification and prediction can facilitate early diagnosis and treatment, improve patients' quality of life, and prevent adverse consequences. Therefore, objective and accurate detection methods are essential. Scalp EEG is inexpensive, convenient, and noninvasive, making it readily applicable in children, adolescents, and clinical patients. Existing research has shown that NSSI is closely associated with impulsive cognitive control. The Go / No-Go task is a classic experimental paradigm commonly used to study response inhibition, impulse control, and attention. In this task, subjects must respond or inhibit a response based on the type of stimulus presented. Previous studies have found that depressed patients with suicidal ideation or behavior exhibit a decrease in the P3 component of the EEG in this task. However, existing technologies have yet to address the core issue of how to construct a recognition model for NSSI using the EEG activity characteristics associated with the Go / No-Go task.

[0003] Existing diagnoses of non-suicidal self-injury (NSSI) rely primarily on subjective clinical interviews and patient self-reports, lacking objective, quantifiable biomarkers. This makes diagnostic results susceptible to individual biases, making it difficult to accurately assess symptom severity, leading to inaccurate and delayed diagnoses. Existing methods based on skin conductivity testing and video behavioral analysis, while partially able to reflect autonomic arousal or external movement characteristics, cannot pinpoint brain dysfunction directly related to NSSI pathology (such as deficits in prefrontal cortex inhibitory control and limbic system emotion regulation). This makes it difficult to effectively identify and monitor NSSI patients early, and thus falls short of meeting clinical needs. Summary of the Invention

[0004] In view of this, it is necessary to provide a method and system for identifying and predicting non-suicidal self-injury, which can provide objective and accurate detection means to make up for the lack of biomarkers in existing technologies, realize early identification and disease monitoring of NSSI, and create favorable conditions for timely intervention and treatment.

[0005] The present invention provides a method for identifying and predicting non-suicidal self-injury, comprising the following steps: a. collecting behavioral data and electroencephalogram (EEG) signals from a subject; b. preprocessing the collected EEG signals; c. analyzing the preprocessed EEG signals; and d. constructing a classification model based on the collected behavioral data and analyzed EEG signals to identify and predict non-suicidal self-injury.

[0006] Preferably, said step a comprises: said behavioral data collection comprises: recording the subject's correct hit rate, false alarm rate, and correct response time, and evaluating the subject's attention control and inhibitory bias to emotional stimuli;

[0007] The EEG signal collection includes: collecting task-state EEG signals and 10-minute resting-state EEG signals of the subject during the period when the subject completes the emotional task.

[0008] Preferably, the step b comprises:

[0009] All EEG signals were filtered to remove high-frequency noise and baseline drift using offline analysis. Bad leads were identified by visual inspection and replaced by interpolation using data from surrounding leads. Finally, the data were re-referenced to the whole-brain average reference.

[0010] For task-state EEG signals: centered on the stimulus event, the signals were segmented into 2.5-second time periods, and independent component analysis was performed to identify and remove eye movements, blinks, and myoelectric artifacts. After segmentation, correction was performed and abnormal segments were removed.

[0011] For resting-state EEG signals: After removing the initial and final 20 seconds of data, the filtered and re-referenced continuous signal was divided into 4-second segments for independent component analysis to identify and remove eye movement, blink, and myoelectric artifact components. The mean of the entire segment was then subtracted to eliminate DC offset, and segments with excessive amplitude were also removed to ensure that the data quality met the requirements of subsequent analysis.

[0012] Preferably, the step c comprises:

[0013] EEG signal analysis includes: event-related potential analysis, time-frequency analysis, and resting-state EEG microstate analysis;

[0014] The event-related potential analysis includes: averaging the EEG segmented data under each condition after preprocessing to obtain an event-related potential waveform;

[0015] The time-frequency analysis involved subtracting the event-related potential component from the single-trial EEG signal to obtain evoked power, calculating spectral power in the 1-40 Hz frequency band using a sliding Hanning window, and normalizing it to the baseline power mean. Subsequently, a cluster permutation test was used to assess power differences in the delta, theta, alpha, and beta frequency bands between non-suicidal self-injurious patients and controls. Finally, the mean power in the theta frequency band at the central midline electrode within the 300-600 ms time window after stimulation was extracted as the core indicator.

[0016] The resting-state EEG microstate analysis includes: bandpass filtering the resting-state EEG signal, identifying the global field power peak and constructing an individual microstate map; after matching the microstate map to a standardized template, back-fitting it to the individual data and extracting parameters for quantifying the dynamic characteristics of the resting-state brain state.

[0017] Preferably, the step d comprises:

[0018] Build a classification model based on behavioral data and EEG signal features;

[0019] Through the forward selection strategy, the classification model is gradually optimized in descending order of feature importance, and the performance is evaluated using three-fold cross-validation, and the classification model with the highest accuracy is finally selected;

[0020] To assess the ability of the classification model to discriminate between patients with non-suicidal self-injury and healthy controls, and between responders and non-responders.

[0021] The present invention provides a non-suicidal self-injury identification and prediction system, which includes an acquisition module, a preprocessing module, an analysis module, and an identification and prediction module, wherein: the acquisition module is used to collect behavioral data and EEG signals from a subject; the preprocessing module is used to preprocess the collected EEG signals; the analysis module is used to analyze the preprocessed EEG signals; and the identification and prediction module is used to construct a classification model based on the collected behavioral data and the analyzed EEG signals, and to identify and predict non-suicidal self-injury.

[0022] Preferably, the acquisition module is specifically used for:

[0023] The behavioral data collection includes: recording the subject's correct hit rate, false alarm rate, and correct response time, and evaluating the subject's attention control and inhibitory bias to emotional stimuli;

[0024] The EEG signal collection includes: collecting task-state EEG signals and 10-minute resting-state EEG signals of the subject during the period when the subject completes the emotional task.

[0025] Preferably, the preprocessing module is specifically used for:

[0026] All EEG signals were filtered to remove high-frequency noise and baseline drift using offline analysis. Bad leads were identified by visual inspection and replaced by interpolation using data from surrounding leads. Finally, the data were re-referenced to the whole-brain average reference.

[0027] For task-state EEG signals: centered on the stimulus event, the signals were segmented into 2.5-second time periods, and independent component analysis was performed to identify and remove eye movements, blinks, and myoelectric artifacts. After segmentation, correction was performed and abnormal segments were removed.

[0028] For resting-state EEG signals: After removing the initial and final 20 seconds of data, the filtered and re-referenced continuous signal was divided into 4-second segments for independent component analysis to identify and remove eye movement, blink, and myoelectric artifact components. The mean of the entire segment was then subtracted to eliminate DC offset, and segments with excessive amplitude were also removed to ensure that the data quality met the requirements of subsequent analysis.

[0029] Preferably, the analysis module is specifically used for:

[0030] EEG signal analysis includes: event-related potential analysis, time-frequency analysis, and resting-state EEG microstate analysis;

[0031] The event-related potential analysis includes: averaging the EEG segmented data under each condition after preprocessing to obtain an event-related potential waveform;

[0032] The time-frequency analysis involved subtracting the event-related potential component from the single-trial EEG signal to obtain evoked power, calculating spectral power in the 1-40 Hz frequency band using a sliding Hanning window, and normalizing it to the baseline power mean. Subsequently, a cluster permutation test was used to assess power differences in the delta, theta, alpha, and beta frequency bands between non-suicidal self-injurious patients and controls. Finally, the mean power in the theta frequency band at the central midline electrode within the 300-600 ms time window after stimulation was extracted as the core indicator.

[0033] The resting-state EEG microstate analysis includes: bandpass filtering the resting-state EEG signal, identifying the global field power peak and constructing an individual microstate map; after matching the microstate map to a standardized template, back-fitting it to the individual data and extracting parameters for quantifying the dynamic characteristics of the resting-state brain state.

[0034] Preferably, the recognition and prediction module is specifically used to:

[0035] Build a classification model based on behavioral data and EEG signal features;

[0036] Through the forward selection strategy, the classification model is gradually optimized in descending order of feature importance, and the performance is evaluated using three-fold cross-validation, and the classification model with the highest accuracy is finally selected;

[0037] To assess the ability of the classification model to discriminate between patients with non-suicidal self-injury and healthy controls, and between responders and non-responders.

[0038] This application integrates multimodal EEG features (task-state P3 amplitude, theta wave power, and resting-state microstate dynamic parameters) to construct an objective detection model based on neural activity characteristics. This allows for accurate early identification of NSSI and dynamic disease monitoring and quantitative assessment, providing a traceable and verifiable basis for clinical diagnosis and treatment. More specifically, the application's beneficial effects include:

[0039] First, precise neural markers: Based on Go / No-Go task-state EEG, the specific neural characteristics of NSSI patients (such as reduced P3 amplitude and attenuated theta wave power) are captured to directly locate the cognitive control deficits in the prefrontal lobe, breaking through the limitations of subjective reports, skin electricity, and video detection that cannot reveal the pathological mechanism.

[0040] Second, dynamic diagnosis and prediction: combining millisecond-level EEG analysis to track the neural dynamics of impulse inhibition in real time, and integrating resting-state microstate characteristics to predict treatment response, to achieve an integrated assessment of "detection-intervention-prognosis".

[0041] Third, efficient clinical application: the device is portable and non-invasive, suitable for adolescents; the algorithm is lightweight, and its computational efficiency is significantly better than video multi-model fusion, making it suitable for clinical real-time screening and long-term monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flow chart of the method for identifying and predicting non-suicidal self-injury of the present invention;

[0043] Figure 2 A schematic diagram of building a classification model based on behavioral data and EEG signal features provided by an embodiment of the present invention;

[0044] Figure 3 FIG. 4 is a hardware architecture diagram of the non-suicidal self-injury identification and prediction system of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] See Figure 1 FIG. 2 is a flowchart of a preferred embodiment of the method for identifying and predicting non-suicidal self-injury according to the present invention.

[0047] Step S1: collect behavioral data and EEG signals from the subject. Specifically:

[0048] Behavioral data collection: This embodiment uses a visual presentation system: it is mainly used to present the visual stimuli of the test, including: a computer equipped with a monitor, and the emotional Go / No-go task presented on the computer screen.

[0049] Grayscale facial images were used as stimuli, encompassing three types of expressions: positive (happy), negative (sad), and neutral. Size, brightness, contrast, and temporal frequency were uniformly adjusted to ensure stimulus standardization. Each face stimulus was presented for 800 milliseconds, followed by a fixation point (lasting 800-1200 milliseconds).

[0050] Participants were required to press a key to respond to a specific type of facial stimulus (Go trials) and not to another type of stimulus (NoGo trials), maximizing speed and accuracy. The task consisted of four conditions, 75% of which were Go trials and 25% were NoGo trials. The stimulus combinations for each group were as follows: happy Go / neutral NoGo, neutral Go / happy NoGo, sad Go / neutral NoGo, and neutral Go / sad NoGo. The order of testing across conditions was randomized and counterbalanced across subjects. The subjects' correct hit rate, false alarm rate, and correct response time were recorded to assess their attentional control and inhibitory bias in response to emotional stimuli.

[0051] EEG signal acquisition: In this example, EEG data were collected using a 32-lead EEG system (sampling rate 500 Hz). The electrode array was arranged based on the extended international 10-20 system, with the online reference electrode at the central midline (Cz) and the ground electrode at the frontal midline (Fz). During recording, electrode impedance was maintained below 10 kilohms (kΩ). Task-state EEG signals were collected while the subjects completed an emotional Go / No-Go task, as well as 10 minutes of resting-state EEG signals.

[0052] Step S2: pre-process the collected EEG signals. Specifically:

[0053] All EEG signals were sequentially filtered using offline analysis to remove high-frequency noise and baseline drift. Bad leads were identified by visual inspection and replaced by interpolation using data from surrounding leads. Finally, the data were rereferenced to the whole-brain average reference.

[0054] For task-phase EEG signals, the signals were segmented into 2.5-second segments (1000 ms before to 1500 ms after the stimulus event) centered on the stimulus event. Independent component analysis (ICA) was performed using the Runica algorithm to identify and remove eye movements, blinks, and myoelectric artifacts. After segmentation, the signals were corrected using the 200 ms before stimulus as the baseline, and outliers with amplitudes exceeding ±100 µV were removed.

[0055] For resting-state EEG signals: After removing the initial and final 20 seconds of data, the filtered and re-referenced continuous signal was segmented into 4-second segments. The ICA artifact removal process was consistent with the task state. The mean was then subtracted from the entire segment to eliminate DC offset, and segments with excessive amplitude were also removed to ensure that the data quality met the requirements of subsequent analysis.

[0056] Step S3: Analyze the pre-processed EEG signal. Specifically:

[0057] EEG signal analysis includes three parts: event-related potential (ERP) analysis, time-frequency analysis, and resting-state EEG microstate analysis.

[0058] For ERP analysis: First, the segmented EEG data under each condition after preprocessing were averaged to obtain the ERP waveform, and the focus was on the P3 component at the Pz electrode point in the time window of 400 to 700 ms after stimulation; the P3 component has been used in previous studies to assess cognitive control and impulsivity in clinical populations and is sensitive to emotional context.

[0059] For time-frequency analysis, evoked power was obtained by subtracting the ERP component from the single-trial EEG signal. Spectral power in the 1-40 Hz band (1 Hz resolution) was calculated using a sliding Hanning window (frequency-dependent time window, 3-18 cycles per frequency) and normalized to the mean power during the baseline period (-200 to 0 ms). Subsequently, a cluster permutation test was used to assess power differences between NSSI patients and controls in the delta (1-3 Hz), theta (4-7 Hz), alpha (8-12 Hz), and beta (13-30 Hz) bands. Finally, the mean power in the theta band at the central midline electrodes (Cz and Fz) within the 300-600 ms post-stimulus time window was extracted as the core measure.

[0060] Resting-state EEG microstate analysis: The resting-state EEG signals were bandpass filtered at 2-20 Hz to identify global field power (GFP) peaks and construct individual microstate maps. First, the individual data of the subjects were clustered at the first level using the k-means algorithm. Then, the subjects in the same group were clustered at the second level (2-8 categories) based on the cross-validation criterion to distinguish four types of EEG microstate maps. After matching the microstate maps to a standardized template, they were back-fitted to the individual data to extract parameters such as the average duration, occurrence frequency, time coverage, and average GFP of each type of microstate to quantify the dynamic characteristics of the resting-state brain state.

[0061] Step S4: Based on the collected behavioral data and the analyzed EEG signals, a classification model is constructed to identify and predict non-suicidal self-injury. Specifically:

[0062] In this embodiment, to achieve the identification of non-suicidal self-injury (NSSI) and prediction of intervention effects, this embodiment constructs a random forest classification model based on behavioral data (such as correct rejection rate and sensitivity) and EEG signal characteristics (task-state P3 amplitude, theta wave power, and resting-state EEG microstate parameters).

[0063] Using the frequency of NSSI at 1, 3, and 6 months post-intervention follow-up as the core predictor, patients were divided into a responder group (no NSSI within 1 month) and a non-responder group. EEG signal features significantly correlated with follow-up indicators were screened. A forward selection strategy was used to optimize the random forest classification model in descending order of feature importance. Performance was evaluated using three-fold cross-validation: the data were split into a training set and a test set at a 2:1 ratio. Validation was repeated until each fold was tested, and the random forest classification model with the highest accuracy was selected. Receiver operating characteristic (ROC) curves were used to calculate the area under the curve (AUC) to evaluate the random forest classification model's ability to discriminate between NSSI patients and healthy controls, and between responders and non-responders.

[0064] It should be noted that the classification model is not limited to the random forest classification model, and can be replaced by a support vector machine (SVM) classification model, a convolutional neural network (CNN) classification model, or other machine learning algorithm classification models.

[0065] See Figure 3 FIG. 1 is a hardware architecture diagram of a non-suicidal self-injury identification and prediction system 10 according to the present invention. The system comprises: an acquisition module 101 , a pre-processing module 102 , an analysis module 103 , and an identification and prediction module 104 .

[0066] The acquisition module 101 is used to collect behavioral data and EEG signals from the subject. Specifically:

[0067] The acquisition module 101 performs behavioral data acquisition: This embodiment adopts a visual presentation system: it is mainly used to present the visual stimulation of the test, including: a computer equipped with a display, and an emotional Go / No-go task presented on the computer screen.

[0068] Grayscale facial images were used as stimuli, encompassing three types of expressions: positive (happy), negative (sad), and neutral. Size, brightness, contrast, and temporal frequency were uniformly adjusted to ensure stimulus standardization. Each face stimulus was presented for 800 milliseconds, followed by a fixation point (lasting 800-1200 milliseconds).

[0069] Participants were required to press a key to respond to a specific type of facial stimulus (Go trials) and not to another type of stimulus (NoGo trials), maximizing speed and accuracy. The task consisted of four conditions, 75% of which were Go trials and 25% were NoGo trials. The stimulus combinations for each group were as follows: happy Go / neutral NoGo, neutral Go / happy NoGo, sad Go / neutral NoGo, and neutral Go / sad NoGo. The order of testing across conditions was randomized and counterbalanced across subjects. The subjects' correct hit rate, false alarm rate, and correct response time were recorded to assess their attentional control and inhibitory bias in response to emotional stimuli.

[0070] The acquisition module 101 acquires EEG signals. In this embodiment, EEG data is collected using a 32-lead EEG system (sampling rate 500 Hz). The electrode array is arranged based on the extended international 10-20 system, with the online reference electrode at the central midline (Cz) and the ground electrode at the frontal midline (Fz). During recording, the electrode impedance is maintained below 10 kilohms (kΩ). Task-state EEG signals are collected while the subject completes an emotional Go / No-Go task, as well as 10 minutes of resting-state EEG signals.

[0071] The pre-processing module 102 is used to pre-process the collected EEG signals. Specifically:

[0072] For all EEG signals, the pre-processing module 102 uses offline analysis to filter sequentially to remove high-frequency noise and baseline drift. Bad leads are identified by visual inspection and interpolated using surrounding lead data. Finally, the data is re-referenced to the whole-brain average reference.

[0073] For task-state EEG signals, the preprocessing module 102 segments the signals into 2.5-second segments (1000 ms before to 1500 ms after the stimulus event) centered on the stimulus event. Independent component analysis (ICA) is performed using the Runica algorithm to identify and remove eye movements, blinks, and myoelectric artifacts. After segmentation, correction is performed using the 200 ms before stimulus as the baseline, and abnormal segments with amplitudes exceeding ±100 µV are removed.

[0074] For resting-state EEG signals: after removing the initial and last 20 seconds of data, the preprocessing module 102 divides the filtered and re-referenced continuous signal into 4-second periods. The ICA artifact removal process is consistent with the task state. Then, the mean of the entire data segment is subtracted to eliminate DC offset, and the periods with excessive amplitude are also eliminated to ensure that the data quality meets the requirements of subsequent analysis.

[0075] The analysis module 103 is used to analyze the pre-processed EEG signal. Specifically:

[0076] EEG signal analysis includes three parts: event-related potential (ERP) analysis, time-frequency analysis, and resting-state EEG microstate analysis.

[0077] The analysis module 103 performs ERP analysis: first, the EEG segmented data under each condition after preprocessing are averaged to obtain an ERP waveform, and the P3 component of the Pz electrode point within the time window of 400 to 700 ms after stimulation is focused on; the P3 component has been used in previous studies to assess cognitive control and impulsivity in clinical populations and is sensitive to emotional context.

[0078] The analysis module 103 performed a time-frequency analysis: evoked power was obtained by subtracting the ERP component from the single-trial EEG signal. Spectral power in the 1-40 Hz frequency band (1 Hz resolution) was calculated using a sliding Hanning window (frequency-dependent time window, 3-18 cycles per frequency) and normalized to the mean power during the baseline period (-200 to 0 ms). Subsequently, a cluster permutation test was used to assess power differences between NSSI patients and controls in the delta (1-3 Hz), theta (4-7 Hz), alpha (8-12 Hz), and beta (13-30 Hz) frequency bands. Finally, the mean power in the theta frequency band at the central midline electrodes (Cz and Fz) within the 300-600 ms post-stimulation time window was extracted as the core indicator.

[0079] The analysis module 103 performs resting-state EEG microstate analysis: 2-20 Hz bandpass filtering is performed on the resting-state EEG signal, the global field power (GFP) peak is identified, and an individual microstate map is constructed: first, the individual data of the subjects are clustered at the first level using the k-means algorithm, and then the subjects in the same group are clustered at the second level (2-8 categories) based on the cross-validation criterion to distinguish four types of EEG microstate maps; after matching the microstate map to a standardized template, it is back-fitted to the individual data to extract parameters such as the average duration, occurrence frequency, time coverage, and average GFP of each type of microstate for quantifying the dynamic characteristics of the resting-state brain state.

[0080] The identification and prediction module 104 is used to construct a classification model based on the collected behavioral data and the analyzed EEG signals, and to identify and predict non-suicidal self-injury. Specifically:

[0081] In this embodiment, to achieve the identification of non-suicidal self-injury (NSSI) and prediction of intervention effects, the identification and prediction module 104 constructs a random forest classification model based on behavioral data (such as correct rejection rate and sensitivity) and EEG signal characteristics (task-state P3 amplitude, theta wave power, and resting-state EEG microstate parameters).

[0082] The identification and prediction module 104 uses the frequency of NSSI at follow-up visits 1, 3, and 6 months after intervention as the core predictive indicator, dividing patients into a responder group (no NSSI within 1 month) and a non-responder group. EEG signal features significantly correlated with these follow-up indicators are screened. A forward selection strategy is used to gradually optimize the random forest classification model in descending order of feature importance. Performance is evaluated using a three-fold cross-validation strategy: the data is split into a training set and a test set at a 2:1 ratio. Validation is repeated until each fold is tested, ultimately selecting the random forest classification model with the highest accuracy. Receiver operating characteristic (ROC) curves are used to calculate the area under the curve (AUC) to evaluate the random forest classification model's ability to discriminate between NSSI patients and healthy controls, and between responders and non-responders.

[0083] It should be noted that the classification model is not limited to the random forest classification model, and can be replaced by a support vector machine (SVM) classification model, a convolutional neural network (CNN) classification model, or other machine learning algorithm classification models.

[0084] Experimental verification:

[0085] Experiment 1: In a Go / No-Go paradigm, positive or negative emotional stimuli were paired with neutral stimuli, with Go trials accounting for 75% of the time. The Go / No-Go role of the emotional stimulus type (positive / negative) and the neutral stimulus was dynamically switched during testing. (See Table 1.) Experiment 1 data further validated that the EEG P3 component and theta wave power attenuation during the task can effectively reveal neurophysiological differences in cognitive control and emotion regulation in patients with NSSI, accurately detecting NSSI.

[0086] Table 1: EEG P3 component and theta wave power attenuation effectively detect NSSI patients

[0087]

[0088] Experiment 2: Combining task-state EEG and resting-state microstate dynamic analysis, we constructed a specific biomarker for NSSI by analyzing the attenuation of P3 amplitude, the reduction of central midline theta wave power, and the change of microstate D coverage, breaking through the limitations of single-modality detection. Please refer to Figure 2 ,in: Figure 2 (A) and (B) are schematic diagrams for distinguishing NSSI patients from healthy control subjects. Figure 2(C) and (D) are schematic diagrams for predicting self-injurious behavior in NSSI patients at the one-month follow-up. Experiment 2 data showed that using the Go condition-theta power, Go condition-P3 amplitude, No-go condition-theta power, and the correct rejection rate of the task effectively identified NSSI patients (AUC = 0.8452). The frequency of resting-state EEG microstate C, the temporal coverage of state D, and the frequency of state D effectively predicted whether patients still engaged in self-injurious behavior at the one-month follow-up (AUC = 0.8362).

[0089] Although the present invention has been described with reference to the current preferred embodiments, those skilled in the art should understand that the above-mentioned preferred embodiments are only used to illustrate the present invention and are not used to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying and predicting non-suicidal self-injury, characterized in that: The method comprises the following steps: a. Collect behavioral data and EEG signals from subjects; b. Preprocessing the collected EEG signals; c. Analyze the preprocessed EEG signals; d. Build a classification model based on collected behavioral data and analyzed EEG signals to identify and predict non-suicidal self-injury; The step a includes: the behavioral data collection includes: recording the subject's correct hit rate, false alarm rate, and correct response time, and evaluating the subject's attention control and inhibition bias to emotional stimuli; the EEG signal collection includes: collecting the subject's task-state EEG signals during the emotional task and 10 minutes of resting-state EEG signals; The step b comprises: for all EEG signals: using offline analysis to filter in sequence to remove high-frequency noise and baseline drift; identifying bad leads by visual inspection, and interpolating and replacing them with surrounding lead data, and finally re-referencing the data to the whole-brain average reference; For task-state EEG signals: centered on the stimulus event, the signals were segmented into 2.5-second time periods, and independent component analysis was performed to identify and remove eye movements, blinks, and myoelectric artifacts. After segmentation, correction was performed and abnormal segments were removed. For resting-state EEG signals: After removing the initial and final 20 seconds of data, the filtered and re-referenced continuous signal was segmented into 4-second segments. Independent component analysis was performed to identify and remove eye movements, blinks, and myoelectric artifacts. The mean of the entire segment was then subtracted to eliminate DC offset. Segments with excessive amplitude were also removed to ensure data quality met the requirements of subsequent analysis. The step c includes: EEG signal analysis including event-related potential analysis, time-frequency analysis and resting-state EEG microstate analysis; The event-related potential analysis includes: averaging the EEG segmented data under each condition after preprocessing to obtain an event-related potential waveform; The time-frequency analysis involved subtracting the event-related potential component from the single-trial EEG signal to obtain evoked power, calculating spectral power in the 1-40 Hz frequency band using a sliding Hanning window, and normalizing it to the baseline power mean. Subsequently, a cluster permutation test was used to assess power differences in the delta, theta, alpha, and beta frequency bands between non-suicidal self-injurious patients and controls. Finally, the mean power in the theta frequency band at the central midline electrode within the 300-600 ms time window after stimulation was extracted as the core indicator. The resting-state EEG microstate analysis includes: bandpass filtering the resting-state EEG signal, identifying the global field power peak and constructing an individual microstate map; matching the microstate map to a standardized template, back-fitting it to the individual data, and extracting parameters for quantifying the dynamic characteristics of the resting-state brain state; The step d includes: constructing a classification model based on behavioral data and EEG signal features; optimizing the classification model step by step in descending order of feature importance through a forward selection strategy, evaluating performance using three-fold cross-validation, and ultimately selecting the classification model with the highest accuracy; and evaluating the classification model's ability to distinguish between non-suicidal self-injury patients and healthy controls, and between responding and non-responding groups.

2. A method for identifying and predicting non-suicidal self-injury according to claim 1, characterized in that: The method also includes a non-suicidal self-injury identification and prediction system, which includes an acquisition module, a pre-processing module, an analysis module, and an identification and prediction module, wherein: The acquisition module is used to collect behavioral data and EEG signals from the subject; The preprocessing module is used to preprocess the collected EEG signals; The analysis module is used to analyze the preprocessed EEG signal; The recognition and prediction module is used to construct a classification model based on the collected behavioral data and the analyzed EEG signals, and to identify and predict non-suicidal self-injury.

3. The method according to claim 2, wherein The acquisition module is specifically used for: The behavioral data collection includes: recording the subject's correct hit rate, false alarm rate, and correct response time, and evaluating the subject's attention control and inhibitory bias to emotional stimuli; The EEG signal collection includes: collecting task-state EEG signals and 10-minute resting-state EEG signals of the subject during the period when the subject completes the emotional task.

4. The method according to claim 3, wherein The preprocessing module is specifically used for: All EEG signals were filtered to remove high-frequency noise and baseline drift using offline analysis. Bad leads were identified by visual inspection and replaced by interpolation using data from surrounding leads. Finally, the data were re-referenced to the whole-brain average reference. For task-state EEG signals: centered on the stimulus event, the signals were segmented into 2.5-second time periods, and independent component analysis was performed to identify and remove eye movements, blinks, and myoelectric artifacts. After segmentation, correction was performed and abnormal segments were removed. For resting-state EEG signals: After removing the initial and final 20 seconds of data, the filtered and re-referenced continuous signal was divided into 4-second segments for independent component analysis to identify and remove eye movement, blink, and myoelectric artifact components. The mean of the entire segment was then subtracted to eliminate DC offset, and segments with excessive amplitude were also removed to ensure that the data quality met the requirements of subsequent analysis.

5. The method according to claim 4, wherein The analysis module is specifically used for: EEG signal analysis includes: event-related potential analysis, time-frequency analysis, and resting-state EEG microstate analysis; The event-related potential analysis includes: averaging the EEG segmented data under each condition after preprocessing to obtain an event-related potential waveform; The time-frequency analysis involved subtracting the event-related potential component from the single-trial EEG signal to obtain evoked power, calculating spectral power in the 1-40 Hz frequency band using a sliding Hanning window, and normalizing it to the baseline power mean. Subsequently, a cluster permutation test was used to assess power differences in the delta, theta, alpha, and beta frequency bands between non-suicidal self-injurious patients and controls. Finally, the mean power in the theta frequency band at the central midline electrode within the 300-600 ms time window after stimulation was extracted as the core indicator. The resting-state EEG microstate analysis includes: bandpass filtering the resting-state EEG signal, identifying the global field power peak and constructing an individual microstate map; after matching the microstate map to a standardized template, back-fitting it to the individual data and extracting parameters for quantifying the dynamic characteristics of the resting-state brain state.

6. The method according to claim 5, wherein The identification and prediction module is specifically used for: Build a classification model based on behavioral data and EEG signal features; Through the forward selection strategy, the classification model is gradually optimized in descending order of feature importance, and the performance is evaluated using three-fold cross-validation, and the classification model with the highest accuracy is finally selected; To assess the ability of the classification model to discriminate between patients with non-suicidal self-injury and healthy controls, and between responders and non-responders.

Images