A virtual simulation training method for psychological intervention of adolescent internet addiction

Abstract

The method for psychological intervention virtual simulation training for adolescent internet addiction belongs to the technical field of digital medical treatment and mental health. The method comprises the following steps: through multi-modal physiological signals and eye movement data, after preprocessing and feature extraction, a craving degree prediction model is trained by using an agent craving label, and attention bias is calculated; a generative adversarial network is driven by a craving degree, an attention bias and a difficulty coefficient to generate a cognitive reevaluation statement and a facial action unit parameter of a virtual avatar in real time; and the difficulty is adjusted based on a frontal lobe theta / beta power ratio closed loop. The present application deeply integrates adaptive generative adversarial intervention and neural closed loop regulation, realizes precise training of cognitive bias and impulsive inhibition, and significantly improves the pertinence and ecological validity of the intervention.

Description

Technical Field

[0001] This invention relates to the fields of digital healthcare and mental health technology, and in particular to a virtual simulation training method for psychological intervention in adolescent internet addiction. Background Technology

[0002] Adolescent internet addiction has become a global public health problem, with its core psychopathological mechanisms manifesting as an abnormal craving for internet-related stimuli, attentional fixation, and impaired impulse control. Virtual reality exposure and biofeedback training, due to their advantages of high immersion and quantifiable feedback, are increasingly being used in addiction intervention. However, existing virtual reality-based psychological intervention techniques for internet addiction still have significant shortcomings.

[0003] First, the generated intervention content is disconnected from the individual's real-time state. Most systems use pre-recorded scripts or fixed virtual character dialogues, failing to perceive the dramatic fluctuations in adolescents' psychological needs during training. When the need level rises sharply, the system still presents mild persuasive content; when the need is low, it may apply excessive stimulation, resulting in poor timeliness and targeting of the intervention. Some methods attempting to quantify need rely on subjective key press reports, interrupting the immersive experience and making it difficult to obtain continuous tags, thus limiting the generation of data-driven adaptive content.

[0004] Second, there is a lack of targeted corrective measures for attentional bias, a key cognitive bias. Current eye-tracking methods in virtual training are mostly used for fixation analysis or interactive control, without transforming attentional bias indicators into dynamic, adversarial cognitive reappraisal training materials. The automatic attentional fixation of adolescents on addictive cues such as games and social network icons has not been directly challenged and reshaped; training remains at the level of passive exposure, lacking proactive intervention in cognitive processing.

[0005] Third, the training difficulty adjustment is crude and fails to form a closed loop with the neurophysiological indicators of impulse inhibition. Some systems adjust scene intensity based on heart rate or skin conductance, but these peripheral physiological indicators cannot specifically reflect the state of prefrontal executive function responsible for impulse control. The power ratio of the prefrontal theta band to the beta band is a recognized neurophysiological marker of cognitive control and impulse inhibition, but no virtual simulation training system currently feeds this indicator back to the generation of intervention strategies and difficulty adjustment in real time, resulting in weak direct training effects on inhibitory control function.

[0006] Fourth, the static nature of the intervention content leads to habituation and difficulty in training transfer. Repeated exposure to the same virtual scenarios and fixed persuasive content causes adolescents to quickly adapt, making it difficult to induce stable cognitive reappraisal abilities and behavioral inhibition skills in real-world internet usage scenarios. Furthermore, the poor physical compatibility and complex wearing of high-density EEG acquisition devices with head-mounted displays further limit the practical application of this type of neural closed-loop training method.

[0007] Therefore, there is an urgent need for a virtual simulation training method that can use lightweight physiological sensing and eye tracking to perceive an individual's craving state and cognitive biases in real time, dynamically generate highly individualized and empathetic adversarial intervention content, and conduct closed-loop regulation based on prefrontal impulse inhibition neural indicators, so as to simultaneously improve adolescents' resistance to internet addiction from both cognitive and neurological levels. Summary of the Invention

[0008] To achieve the above objectives, this invention provides a virtual simulation training method for psychological intervention in adolescent internet addiction, comprising the following steps:

[0009] Step 1: Synchronously collect the user's multimodal physiological signals and eye movement data. The multimodal physiological signals include electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, and skin conductance signals. The eye movement data includes fixation point coordinates.

[0010] Step 2: Preprocess and extract features from the electroencephalogram (EEG), electrocardiogram (ECG), skin conductance signals, and eye movement data to form a multidimensional feature vector sequence;

[0011] Step 3: Input the multidimensional feature vector sequence into a pre-trained thirst prediction model based on a long short-term memory network and a scaled dot product attention mechanism, and output the thirst index at the current time. The thirst prediction model uses surrogate thirst labels for supervised learning during the training phase. The surrogate thirst labels are generated from peripheral physiological signals through a calibration mapping model.

[0012] Step 4: Using the fixation point coordinates in the eye-tracking data, dynamically update the region of interest (ROI) using a raycasting algorithm within the virtual scene, and calculate the proportion of fixation time on addiction-related ROIs within a sliding time window to construct a cognitive bias vector. ;

[0013] Step 5: Using the aforementioned desire index The cognitive bias vector and current difficulty level As a condition, an input condition generative adversarial network is generated to generate an interactive behavior strategy for the virtual avatar. The interactive behavior strategy includes cognitive re-evaluation guidance statements and facial action unit intensity parameters.

[0014] Step 6: In the virtual scene presented by the head-mounted virtual reality display, render a virtual clone that acts according to the interaction behavior strategy, and the virtual clone performs cognitive reappraisal guidance and impulse inhibition tasks;

[0015] Step 7: During the interaction, continuously collect EEG signals from the prefrontal cortex electrodes and calculate the ratio of theta band power to beta band power in real time as an indicator of impulse inhibition. ;

[0016] Step 8: Compare the impulse inhibition indices With pre-established individualization thresholds Update the difficulty coefficient based on the comparison results. And the updated difficulty level Feedback is sent to step 5 for the generation of the next round of interaction behavior strategies;

[0017] Step 9: Repeat steps 5 to 8 until the training duration for a single session reaches the preset value.

[0018] Preferably, in step 1, the electroencephalogram (EEG) signal is acquired using a forehead dry electrode headband at a sampling rate of 250 Hz. The forehead dry electrode headband contains Fp1, Fp2, and Fz electrodes placed according to the international 10-20 system. The electrocardiogram (ECG) signal is acquired using a chest strap heart rate sensor. The skin conductance signal is acquired using a finger electrode skin conductance sensor. The gaze point coordinates and pupil diameter data are acquired using an infrared eye tracker built into a head-mounted virtual reality display.

[0019] Preferably, in step 2, the preprocessing includes: performing bandpass filtering on the EEG signal from 0.5 Hz to 45 Hz, and using a pre-trained single-channel deep denoising network to remove eye movement artifacts from the EEG signal of each electrode, wherein the single-channel deep denoising network is a convolutional autoencoder trained on the synthesized clean EEG signal and the noisy EEG signal; extracting the RR interval sequence from the ECG signal; and separating the skin conductance level component and the skin conductance response component from the skin conductance signal.

[0020] The feature extraction includes: calculating the power spectral density of the theta band (4 Hz to 8 Hz) and the power spectral density of the beta band (13 Hz to 30 Hz) using short-time Fourier transform; calculating the heart rate variability time-domain index RMSSD and the frequency-domain index low-frequency power to high-frequency power ratio; extracting the average skin conductance and skin conductance response amplitude; extracting fixation duration, saccade amplitude, and pupil diameter variability from the eye movement data; and aligning all features with a 100-millisecond time window to form the multidimensional feature vector sequence.

[0021] Preferably, in step 3, the thirst prediction model based on long short-term memory network and scaled dot product attention mechanism encodes the multidimensional feature vector sequence using a two-layer unidirectional long short-term memory network, outputting a hidden state sequence. Then, the context vector is calculated via a scaled dot product attention mechanism. After concatenation, the desire index is output through a fully connected layer and a sigmoid function. :

[0022]

[0023] in, It is in a hidden state. For context vectors, For the weight vector, For bias, It is an S-shaped function, and its output value is... ;

[0024] The calibration mapping model collects subjective visual analog scale scores provided by adolescent subjects in real time under simulated network craving conditions offline, using subjective scores as the target variable and skin conductance response amplitude as the target variable. Heart rate variability, low-frequency to high-frequency power ratio And pupil diameter variability To predict the variables, train a ridge regression mapping model, with regularization coefficients... Determined through 5-fold cross-validation; in application, each time window's... , and Input the ridge regression mapping model and output the values ​​mapped to the [0,1] interval by the sigmoid function, which are used as the proxy desired labels.

[0025] Preferably, in step 4, the areas of interest include online game icon areas of interest, social media icon areas of interest, and neutral object areas of interest; the raycasting algorithm detects the collision between the gaze point coordinates and the virtual object in real time, dynamically determining the boundaries of each type of area of ​​interest; the sliding time window is 3 seconds long, and the proportion of the user's total gaze time on the online game icon area of ​​interest within the sliding time window to the total gaze time of all areas of interest is calculated to obtain the game attention bias value. The social attention bias value is obtained by measuring the percentage of total gaze time on the interest areas of the aforementioned social media icons. ,calculate and The average value is This constitutes a 3-dimensional cognitive bias vector. .

[0026] Preferably, in step 5, the desire index is... The cognitive bias vector and difficulty level Concatenate into a condition vector ; the condition vector The mapping is a continuous prompt embedding, which is injected into the self-attention layer of the pre-trained language model. The pre-trained language model is a lightweight large language model obtained through knowledge distillation, which is deployed on a local edge computing unit and quantized with INT8. The cognitive reassessment guidance statement sequence is generated by autoregression.

[0027] The facial motion unit intensity parameter sequence is generated by a conditional variational autoencoder, the conditional input of which includes the conditional vector. And the average hidden state vector of the last layer decoder when the lightweight large language model generates a sentence sequence;

[0028] The discriminator of the conditional generative adversarial network is a convolutional neural network discriminator, which uses the virtual clone behavior sequence and the conditional vector. The input is a conditional matching score, and the output is subject to cross-modal consistency loss penalty during training. The cross-modal consistency loss increases when the input facial action unit intensity parameter sequence does not match the emotional polarity carried by the cognitive reappraisal guidance statement sequence.

[0029] Preferably, in step 6, the virtual scene is controlled by a scene manager based on a finite state machine. The finite state machine includes an initial induced state, an adversarial debate state, and a behavior inhibition task state. The transitions between states are achieved using easing functions to gradually change scene elements over 2 to 3 seconds. In the adversarial debate state, the scene manager controls the scene based on the desire index. The virtual avatar dynamically combines elements of the online game interface, short video waterfall, and social message pop-up with the user's historical exposure records. The virtual avatar issues the cognitive reassessment guidance statement for 30 to 60 seconds. Then it switches to the behavior inhibition task state, where the virtual avatar displays an online temptation icon and asks the user to perform the Go / No-go inhibition button operation for 15 to 30 seconds. The two states alternate and cycle until the total training time for a single session reaches 20 minutes.

[0030] Preferably, in step 7, the acquired EEG signals from the prefrontal electrodes are bandpass filtered from 0.5 Hz to 45 Hz, and the total power of the theta band from 4 Hz to 8 Hz is calculated in real time using short-time Fourier transform. The total power of the beta band from 13 Hz to 30 Hz The impulse inhibition index The calculation is as follows:

[0031]

[0032] A higher value indicates a weaker impulse inhibition ability.

[0033] Preferably, in step 8, the individualized threshold... A 5-minute resting-state baseline was collected before the first training session. After removing the data from the first minute of the adaptation period, the mean theta / beta power ratio of the prefrontal cortex was calculated for the remaining 4 minutes. The difficulty coefficient... The update rules are as follows:

[0034]

[0035] in, To adaptively adjust the step size, , Impulse inhibition index for the past 5 iterations With the individualized threshold The root mean square error, The function will adjust the difficulty level. Limited to Interval; when At that time, reduce the difficulty level To reduce the intensity of the confrontation and the attractiveness of the seductive scene of the virtual clone; when At that time, increase the difficulty level To enhance the cognitive challenge and salience of the enticing stimulation of virtual avatars.

[0036] Preferably, in step 9, the preset time is 20 minutes. After training, a training report is output. The training report includes: a thirst reduction rate obtained by comparing the mean thirst index of the last 5 minutes before the end of training with the baseline mean thirst collected in a neutral scenario before training; a cognitive bias improvement rate obtained by comparing the mean cognitive bias vector of the last 5 minutes with the mean cognitive bias vector of the initial 5 minutes; and an impulse inhibition index of the last 5 minutes. Mean relative to the individualized threshold The percentage improvement; the inhibition success rate sequence for all behavioral inhibition task stages; and the natural language summary generated based on the above indicators; all raw EEG signals, ECG signals, skin conductance signals and eye movement data are discarded after feature extraction is completed on the local edge device, and only the desensitized statistical indicators and the training report are retained.

[0037] The beneficial effects of this invention are:

[0038] The method provided by this invention produces a synergistic effect that significantly exceeds the simple summation of individual technical features. Specifically, it utilizes a lightweight dry electrode headband to replace the multi-lead wet electrode cap and solves the eye movement artifact problem without a reference channel through a single-channel deep denoising network, greatly improving wearing comfort and system compatibility while ensuring reliable acquisition of the theta / beta ratio in the prefrontal cortex. By collecting a small number of subjective ratings offline to establish a calibration mapping model, skin conductance, heart rate variability, and pupil diameter variability are fused into a high-confidence proxy thirst label, overcoming the difficulty of real-time annotation and enabling the LSTM attention model to output an accurate thirst index. Based on this thirst index and dynamic attention bias vector, a lightweight large language model that has been knowledge distilled and quantified is driven to jointly generate highly empathetic and emotionally consistent cognitive re-evaluation dialogues and behaviors with a facial variational autoencoder, and the cross-modal consistency penalty of the discriminator ensures the matching of facial expressions and semantics. Furthermore, the theta / beta ratio of the prefrontal cortex is introduced as a closed-loop feedback mechanism for impulse inhibition. Combined with individualized baselines, amplitude-limited adaptive step sizes, and alternating dialogue challenges and behavioral inhibition phases, this approach avoids cognitive overload while ensuring that the intervention load consistently matches the executive function state, specifically strengthening the prefrontal cortex's top-down control over craving impulses. The inclusion of a dynamic, gradual scenario state machine and comprehensive efficacy reports enhances the interpretability and clinical applicability of long-term training while preventing habituation. The overall approach constitutes a creative and strong coupling of "decoding psychological states → generative adversarial intervention → neural closed-loop regulation," significantly improving the ecological validity, transferability, and long-term effects of interventions for adolescent internet addiction. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of the steps of the method of the present invention. Detailed Implementation

[0041] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0042] Please see Figure 1This invention provides a virtual simulation training method for psychological intervention in adolescent internet addiction. It integrates virtual reality exposure, multimodal physiological signal processing, deep attention mechanisms, conditional generative adversarial networks, and prefrontal feedback loops into a synergistic system to simultaneously correct adolescents' craving for online stimuli, attentional bias, and impulsive inhibition deficiencies. The following detailed descriptions through embodiments provide explanations for key terms, parameters, and creative aspects that may be unclear.

[0043] In this invention, "virtual avatar" refers to a humanoid 3D character rendered through a head-mounted virtual reality display. Its behavior is dynamically generated by the system based on the user's real-time state, including verbal dialogue and facial expressions. "Attentional bias" refers to the automatic preferential fixation tendency shown by adolescents towards addictive visual stimuli such as online game icons and social media icons, quantified by the proportion of eye-tracking fixation time. "Impulse inhibition index" specifically refers to the ratio of theta-band power to beta-band power recorded by prefrontal cortex electrodes. This ratio is a recognized neurophysiological marker of cognitive control and impulse inhibition ability; a higher value indicates weaker inhibitory control.

[0044] The "surrogate thirst label" in step 3 is an alternative designed because subjective thirst ratings are difficult to collect continuously in real time during training. Its core innovation lies in: offline collection of subjective visual analog scale ratings from a certain number of subjects, establishing a ridge regression mapping model from skin conductance response amplitude, low-frequency to high-frequency power ratio of heart rate variability, and pupil diameter variability to subjective thirst. In the online phase, this model generates continuous surrogate labels highly correlated with subjective ratings for each time window, used to train the thirst prediction model. Regularization coefficient. Five-fold cross-validation was used to determine the model and avoid overfitting. Offline participants included individuals with varying degrees of addiction (internet addiction scale scores ranging from 42 to 78, covering mild to severe) and different genders to ensure the universality of the mapping model. Although there was a time lag between offline data collection and online application, related studies have shown that the mapping relationship between peripheral physiological indicators and subjective cravings remains stable over several weeks, allowing for periodic recalibration before practical application.

[0045] The conditional generative adversarial network in step 5 integrates a lightweight large language model with a multimodal generative architecture based on a conditional variational autoencoder. The language model in the generator is obtained through knowledge distillation using a T5-small architecture, while the teacher model is a GPT-4o-mini. The distillation dataset contains 20,000 cognitive reappraisal dialogue samples generated by the teacher model under given addictive psychological states. Each sample contains 50 to 120 tokens. The loss function uses a weighted sum of cross-entropy and KL divergence, and the distillation temperature is set to 3.0 to ensure that the student model can still generate high-quality empathic statements with 60M parameters. This model is deployed on a local edge computing unit after INT8 quantization, with an autoregressive inference latency of approximately 120 milliseconds. The facial action unit intensity parameters are generated by the conditional variational autoencoder, whose conditional input includes not only conditional vectors... Furthermore, the hidden state average vector of the last layer of the language model decoder is concatenated to inject textual sentiment information into the face generation process. The discriminator additionally introduces a cross-modal consistency loss: DistilBERT is used as the text sentiment recognition model, and AU-EmotionNet is used as the sentiment recognition model based on facial action units. The sentiment polarity vectors are output separately, and their cosine similarity is calculated. When the similarity is below 0.7, a penalty term proportional to the difference is applied. The threshold of 0.7 is determined empirically based on public datasets, effectively suppressing inconsistent sentiment generation samples.

[0046] The adaptive adjustment of the difficulty coefficient in step 8 avoids oscillation by limiting the step size. According to the formula... ,when hour It reaches the minimum value of 0.1; when hour It reached the upper limit of 0.3. Therefore The actual range of values ​​is It balances response speed and stability. Individualized threshold. The 5-minute resting state data collection requires the user to close their eyes and relax, and the data from the first minute is discarded to eliminate the initial adjustment effect.

[0047] Step 6 employs a finite state machine to control the switching between three states of the virtual scene, and uses a easing function to gradually change scene elements, preventing motion sickness and cognitive interruption caused by sudden changes. The dialogue challenge phase and the behavior inhibition task phase alternate to avoid cognitive overload caused by simultaneously applying auditory and visual instructions, ensuring in-depth processing of cognitive reassessment content. It is important to note that the baseline thirst acquisition before training is independent of the finite state machine operation: before the state machine starts, a 60-second independent neutral scene is presented as a baseline acquisition phase; after this phase, the state machine officially starts and enters the initial evoked state, thereby ensuring the purity of the baseline data.

[0048] Example 1

[0049] Taking a 15-year-old internet gaming addict undergoing a single 20-minute virtual simulation training session as an example, the implementation details are explained in detail.

[0050] Step 1: Simultaneously acquire multimodal physiological signals and eye movement data.

[0051] The user wears a forehead dry electrode headband with three built-in dry electrodes (Fp1, Fp2, and Fz), maintaining a contact impedance below 50 kΩ, and recording EEG signals at a sampling rate of 250 Hz. Simultaneously, a chest strap-type heart rate sensor is worn, acquiring ECG signals at 1000 Hz; a finger electrode skin conductivity sensor is held between the left index finger and the eye, acquiring skin conductance signals at 10 Hz. The head-mounted virtual reality display has a 110-degree field of view and a monocular resolution of 2160×2160, and includes an infrared eye tracker that simultaneously acquires binocular gaze coordinates and pupil diameter data at 120 Hz. All sensors transmit timestamped data streams to the local edge computing unit via Bluetooth 5.0 or USB.

[0052] Step 2: Preprocessing and feature extraction to form a multidimensional feature vector sequence.

[0053] Three EEG signals were subjected to 4th-order Butterworth bandpass filtering from 0.5 Hz to 45 Hz, and then fed into a single-channel deep denoising network to remove eye movement artifacts. This denoising network was a convolutional autoencoder, with the encoder containing 5 layers of one-dimensional convolutions (3 kernels, 2 strides, and 16, 32, 64, 128, and 256 channels respectively), and the decoder was a symmetric transposed convolution. Training data was generated as follows: clean resting-state EEG data (closed-eye relaxation state, 5 minutes per segment) from 30 subjects were selected from a publicly available EEG dataset; 2-second segments were randomly extracted from the EOG artifact signals collected from another 10 subjects performing an active saccade task in a VR environment. These were then superimposed onto the clean EEG data at 5 different signal-to-noise ratios to construct 30,000 training samples, with the mean squared error as the loss function. After denoising, a short-time Fourier transform was performed using a 1-second Hanning window and a 0.5-second overlap to calculate the theta (4-8 Hz) and beta (13-30 Hz) power spectral densities. The ECG signal was analyzed using the Pan-Tompkins algorithm to detect the R-wave, obtaining the RR interval sequence, and calculating the RMSSD and the ratio of low-frequency power (0.04-0.15 Hz) to high-frequency power (0.15-0.4 Hz). Skin conductance signals were separated into horizontal and reactive components using a first-order filter with a 0.05 Hz critical frequency, and the mean skin conductance level and peak-to-peak skin conductance response were extracted. For eye-tracking data, fixation and saccades were defined using a velocity threshold of 30 degrees / second, and fixation duration, saccade amplitude, and pupil diameter variability (standard deviation) were extracted. All features were aligned with a 100-millisecond time window to form a 28-dimensional feature vector sequence. .

[0054] Step 3: Calculate the Desire Index .

[0055] A 2-layer unidirectional LSTM (128 hidden state dimensions) encodes the feature sequence, and scales the dot product attention ( ) Calculate the context vector After concatenation, the output is generated by a fully connected layer and a sigmoid function. :

[0056]

[0057] During the training phase, 30 adolescents aged 13-18 with internet addiction (18 males and 12 females, with internet addiction scale scores of 42-78, covering mild to moderate cases) were recruited offline. Visual Analogue Scale (VAS) scores (0-100 points, normalized to [0,1]) were recorded every 30 seconds in virtual temptation scenarios. Simultaneously, [the data was also recorded]. , and The training and validation sets were divided in an 8:2 ratio, and 5-fold cross-validation was used to determine the ridge regression regularization coefficients to obtain the optimal results. When training the LSTM-Attention model, the proxy labels are used as the target, mean squared error loss is used, Adam optimizer is used, learning rate is 0.001, batch size is 64, and training is performed for 300 epochs until the validation set loss converges.

[0058] Step 4: Construct a cognitive bias vector .

[0059] The virtual scene is rendered using the Unity engine, with pre-defined colliders for three types of interest areas (online game icons, social media icons, and neutral objects). Each frame uses Physics.Raycast to detect gaze point hits. The percentage of gaze duration is calculated using a 3-second sliding time window. (Note the bias value in the game) (Social attention bias) and (The average of both) constitutes .

[0060] Step 5: Generate interaction behavior strategy.

[0061] Concatenate condition vectors (5-dimensional), mapped to three consecutive cue embeddings via a linear layer, and injected into the first three self-attention layers of a distilled lightweight T5 model. This model autoregressively generates cognitive reappraisal statements with up to 50 tokens. Simultaneously, a conditional variational autoencoder receives... The input is concatenated with the average vector of the last hidden state of the language model decoder to generate 18-dimensional facial action unit intensity parameters for the next 2 seconds (30 frames / second). The discriminator is a 3-layer one-dimensional convolutional network, jointly trained with WGAN-GP loss and cross-modal consistency loss. The cross-modal consistency loss is linearly penalized when the cosine similarity between the emotion polarity vectors of DistilBERT and AU-EmotionNet is less than 0.7.

[0062] Step 6: Virtual scene presentation and interaction.

[0063] After training begins, users first enter a 60-second isolated neutral scenario phase. This phase presents a vast virtual grassland without any network-related stimuli, used to stabilize the user's state and collect baseline thirst averages. After the 60-second baseline phase, the finite state machine officially starts, entering the initial induced state. In the initial induced state, the virtual scene presents a static, large-scale multiplayer online game wallpaper as a low-intensity enticement element for 5 seconds, smoothly transitioning to the adversarial debate state within 2 seconds using an easing function. In the adversarial debate state, the scene manager... Dynamically combined elements: when When in use, a dynamic game interface and a short video waterfall flow are displayed; otherwise, social message pop-ups and light game icons are displayed. The virtual avatar reads cognitive reassessment statements in the tone of a peer using a TTS engine, and facial expressions are driven by blendshape. After the adversarial debate state lasts for 30-60 seconds (dynamically determined according to the boundaries of sentence paragraphs), the system switches to the behavior inhibition task state. In the behavior inhibition task state, the virtual avatar becomes a task instructor, presenting 20 online temptation icons and 20 neutral icons in random order (icon size 6 degrees of viewing angle, centered position, 2-second interval). The user performs button suppression on the temptation icons, and the success rate of correct suppression is recorded. After the behavior inhibition task state lasts for 15-30 seconds, it returns to the adversarial debate state, and this alternation cycle continues until the total training time reaches 20 minutes.

[0064] Step 7: Calculate the impulse inhibition index .

[0065] EEG signals from Fp1, Fp2, and Fz electrodes were continuously acquired, bandpass filtered from 0.5 to 45 Hz, and short-time Fourier transforms were performed with a 1-second sliding window and a 0.5-second step size. and ,get Updated every 0.5 seconds.

[0066] Step 8: Update the difficulty level .

[0067] Before the user's first training session, they were asked to relax with their eyes closed for 5 minutes in a quiet environment. Prefrontal cortex EEG was recorded, and the mean theta / beta ratio was calculated after discarding the first minute of data as an individualized threshold. During training, after each dialogue-inhibition loop (approximately 90 seconds), the most recent 5 dialogues are calculated. and RMSE, according to Calculate the step size. The range of values ​​is Update according to the rules And feedback is sent to step 5. If but Lowering the level of challenge reduces its intensity; conversely, increasing it enhances cognitive resistance.

[0068] Step 9: Repeat steps 5-8 for up to 20 minutes to generate a report.

[0069] At the end of training, the baseline mean thirst recorded during the independent neutral scene phase (first 60 seconds) is automatically compared with the mean thirst recorded during the last 5 minutes. The mean was used to calculate the rate of decrease in desire; the initial 5 minutes were compared with the final 5 minutes. Calculate the improvement rate using the mean; calculate relative mean The percentage improvement was calculated; the success rate sequence of Go / No-go suppression was statistically analyzed; and a natural language summary was generated. All raw data was discarded after feature extraction; only anonymized metrics were uploaded to ensure privacy.

[0070] Comparative Example

[0071] A fixed-script virtual training method was used for comparison. The fixed-script method uses the same hardware and virtual environment, and also includes a 60-second independent neutral scene baseline phase and subsequent state machine structure. However, the virtual avatar dialogue consists of four pre-recorded audio segments (contents are "Use the internet in moderation," "Participate in more outdoor activities," "Think about your goals," and "Communicate more with your family"), which are played in a loop regardless of the user's state; the scene elements are fixed, with no dynamic combinations; and the difficulty level is [not specified]. The value was fixed at 0.5; there was no closed-loop theta / beta regulation in the prefrontal cortex. The presentation order and duration of the Go / No-go tasks were completely consistent with the previous example. Forty adolescents aged 13-18 years who met the diagnosis of the Internet Addiction Scale were randomly assigned to the experimental group and the control group, with 20 participants in each group. Training was conducted 3 times a week for 20 minutes each time, for 4 weeks. A mixed-effects model was used for analysis. The results are shown in Table 1.

[0072] Table 1. Comparison of the results of the method of the present invention and the fixed script method after 4 weeks of training.

[0073] Therapeutic efficacy indicators (mean ± standard deviation) Method of the present invention Fixed script method Demand decrease rate (%) 43.7±9.1 17.2±11.3 Note the bias towards improvement rate (%) 38.5±8.7 10.8±7.5 Percentage improvement in impulse inhibition indicators (%) 29.4±7.2 5.9±5.1 Increased success rate of Go / No-go task suppression (%) 34.1±10.5 8.3±8.9 Self-assessed reduction rate of internet usage time (%) 36.2±13.4 11.9±14.2

[0074] As shown in Table 1, the method of the present invention significantly outperforms the fixed-script method in all core indicators, confirming the synergistic effectiveness of closed-loop neural modulation and generative adversarial intervention. The above embodiments are merely examples, and any equivalent modifications based on the spirit of the present invention should fall within the scope of protection.

[0075] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0076] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A virtual simulation training method for psychological intervention in adolescent internet addiction, characterized in that: Includes the following steps: Step 1: Synchronously collect the user's multimodal physiological signals and eye movement data. The multimodal physiological signals include electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, and skin conductance signals. The eye movement data includes fixation point coordinates. Step 2: Preprocess and extract features from the electroencephalogram (EEG), electrocardiogram (ECG), skin conductance signals, and eye movement data to form a multidimensional feature vector sequence; Step 3: Input the multidimensional feature vector sequence into a pre-trained thirst prediction model based on a long short-term memory network and a scaled dot product attention mechanism, and output the thirst index at the current time. The thirst prediction model uses surrogate thirst labels for supervised learning during the training phase. The surrogate thirst labels are generated from peripheral physiological signals through a calibration mapping model. Step 4: Using the fixation point coordinates in the eye-tracking data, dynamically update the region of interest (ROI) using a raycasting algorithm within the virtual scene, and calculate the proportion of fixation time on addiction-related ROIs within a sliding time window to construct a cognitive bias vector. ; Step 5: Using the aforementioned desire index The cognitive bias vector and current difficulty level As a condition, an input condition generative adversarial network is generated to generate an interactive behavior strategy for the virtual avatar. The interactive behavior strategy includes cognitive re-evaluation guidance statements and facial action unit intensity parameters. Step 6: In the virtual scene presented by the head-mounted virtual reality display, render a virtual clone that acts according to the interaction behavior strategy, and the virtual clone performs cognitive reappraisal guidance and impulse inhibition tasks; Step 7: During the interaction, continuously collect EEG signals from the prefrontal cortex electrodes and calculate the ratio of theta band power to beta band power in real time as an indicator of impulse inhibition. ; Step 8: Compare the impulse inhibition indices With pre-established individualization thresholds Update the difficulty coefficient based on the comparison results. And the updated difficulty level Feedback is sent to step 5 for the generation of the next round of interaction behavior strategies; Step 9: Repeat steps 5 to 8 until the training duration for a single session reaches the preset value.

2. The virtual simulation training method for psychological intervention in adolescent internet addiction according to claim 1, characterized in that, In step 1, the electroencephalogram (EEG) signal is acquired at a sampling rate of 250 Hz using a frontal dry electrode headband. The frontal dry electrode headband has Fp1, Fp2, and Fz electrodes placed according to the international 10-20 system. The electrocardiogram (ECG) signal is acquired using a chest strap heart rate sensor. The skin conductance signal is acquired using a finger electrode skin conductance sensor. The gaze point coordinates and pupil diameter data are acquired using an infrared eye tracker built into a head-mounted virtual reality display.

3. The virtual simulation training method for psychological intervention in adolescent internet addiction according to claim 2, characterized in that, In step 2, the preprocessing includes: performing bandpass filtering on the EEG signal from 0.5 Hz to 45 Hz, and using a pre-trained single-channel deep denoising network to remove eye movement artifacts from the EEG signal of each electrode. The single-channel deep denoising network is a convolutional autoencoder trained on the synthesized clean EEG signal and the noisy EEG signal; extracting the RR interval sequence from the ECG signal; and separating the skin conductance level component and the skin conductance response component from the skin conductance signal. The feature extraction includes: calculating the power spectral density of the theta band (4 Hz to 8 Hz) and the power spectral density of the beta band (13 Hz to 30 Hz) using short-time Fourier transform; calculating the heart rate variability time-domain index RMSSD and the frequency-domain index low-frequency power to high-frequency power ratio; extracting the average skin conductance and skin conductance response amplitude; extracting fixation duration, saccade amplitude, and pupil diameter variability from the eye movement data; and aligning all features with a 100-millisecond time window to form the multidimensional feature vector sequence.

4. The virtual simulation training method for psychological intervention in adolescent internet addiction according to claim 3, characterized in that, In step 3, the thirst prediction model based on long short-term memory network and scaled dot product attention mechanism encodes the multidimensional feature vector sequence using a two-layer unidirectional long short-term memory network, outputting a hidden state sequence. Then, the context vector is calculated via a scaled dot product attention mechanism. After concatenation, the desire index is output through a fully connected layer and a sigmoid function. : in, It is in a hidden state. For context vectors, For the weight vector, For bias, It is an S-shaped function, and its output value is... ; The calibration mapping model collects subjective visual analog scale scores provided by adolescent subjects in real time under simulated network craving conditions offline, using subjective scores as the target variable and skin conductance response amplitude as the target variable. Heart rate variability, low-frequency to high-frequency power ratio And pupil diameter variability To predict the variables, train a ridge regression mapping model, with regularization coefficients... Determined through 5-fold cross-validation; in application, each time window's... , and Input the ridge regression mapping model and output the values ​​mapped to the [0,1] interval by the sigmoid function, which are used as the proxy desired labels.

5. The virtual simulation training method for psychological intervention in adolescent internet addiction according to claim 4, characterized in that, In step 4, the areas of interest include online game icon areas of interest, social media icon areas of interest, and neutral object areas of interest; the raycasting algorithm detects the collision between the gaze point coordinates and virtual objects in real time, dynamically determining the boundaries of each type of area of ​​interest; the sliding time window is 3 seconds long, and the proportion of the total gaze time of the user on the online game icon area of ​​interest within the sliding time window to the total gaze time of all areas of interest is calculated to obtain the game attention bias value. The social attention bias value is obtained by measuring the percentage of total gaze time on the interest areas of the aforementioned social media icons. ,calculate and The average value is This constitutes a 3D cognitive bias vector. .

6. The virtual simulation training method for psychological intervention in adolescent internet addiction according to claim 5, characterized in that, In step 5, the desire index is... The cognitive bias vector and difficulty level Concatenate into a condition vector ; the condition vector The mapping is a continuous prompt embedding, which is injected into the self-attention layer of the pre-trained language model. The pre-trained language model is a lightweight large language model obtained through knowledge distillation, which is deployed on a local edge computing unit and quantized with INT8. The cognitive reassessment guidance statement sequence is generated by autoregression. The facial motion unit intensity parameter sequence is generated by a conditional variational autoencoder, the conditional input of which includes the conditional vector. And the average hidden state vector of the last layer decoder when the lightweight large language model generates a sentence sequence; The discriminator of the conditional generative adversarial network is a convolutional neural network discriminator, which uses the virtual clone behavior sequence and the conditional vector. The input is a conditional matching score, and the output is subject to cross-modal consistency loss penalty during training. The cross-modal consistency loss increases when the input facial action unit intensity parameter sequence does not match the emotional polarity carried by the cognitive reappraisal guidance statement sequence.

7. The virtual simulation training method for psychological intervention in adolescent internet addiction according to claim 6, characterized in that, In step 6, the virtual scene is controlled by a scene manager based on a finite state machine. This finite state machine includes an initial induced state, an adversarial debate state, and a behavior inhibition task state. The transitions between states are achieved using easing functions to gradually change scene elements over 2 to 3 seconds. In the adversarial debate state, the scene manager controls the scene based on the desiredness index. The virtual avatar dynamically combines elements of the online game interface, short video waterfall, and social message pop-up with the user's historical exposure records. The virtual avatar issues the cognitive reassessment guidance statement for 30 to 60 seconds. Then it switches to the behavior inhibition task state, where the virtual avatar displays an online temptation icon and asks the user to perform the Go / No-go inhibition button operation for 15 to 30 seconds. The two states alternate and cycle until the total training time for a single session reaches 20 minutes.

8. The virtual simulation training method for psychological intervention in adolescent internet addiction according to claim 7, characterized in that, In step 7, the acquired EEG signals from the prefrontal electrodes are bandpass filtered from 0.5 Hz to 45 Hz, and the total power of the theta band from 4 Hz to 8 Hz is calculated in real time using short-time Fourier transform. The total power of the beta band from 13 Hz to 30 Hz The impulse inhibition index The calculation is as follows: A higher value indicates a weaker impulse inhibition ability.

9. The virtual simulation training method for psychological intervention in adolescent internet addiction according to claim 8, characterized in that, In step 8, the individualized threshold A 5-minute resting-state baseline was collected before the first training session. After removing the data from the first minute of the adaptation period, the mean theta / beta power ratio of the prefrontal cortex was calculated for the remaining 4 minutes. The difficulty coefficient... The update rules are as follows: in, To adaptively adjust the step size, , Impulse inhibition index for the past 5 iterations With the individualized threshold The root mean square error, The function will adjust the difficulty level. Limited to Interval; when At that time, reduce the difficulty level To reduce the intensity of the confrontation and the attractiveness of the seductive scene of the virtual clone; when At that time, increase the difficulty level To enhance the cognitive challenge and salience of the enticing stimulation of virtual avatars.

10. The virtual simulation training method for psychological intervention in adolescent internet addiction according to claim 9, characterized in that, In step 9, the preset time is 20 minutes. After training, a training report is output. The training report includes: the thirst reduction rate obtained by comparing the mean thirst index in the last 5 minutes before the end of training with the baseline mean thirst collected in a neutral scenario before training; the cognitive bias improvement rate obtained by comparing the mean cognitive bias vector in the last 5 minutes with the mean cognitive bias vector in the initial 5 minutes; and the impulse inhibition index in the last 5 minutes. Mean relative to the individualized threshold The percentage improvement; the inhibition success rate sequence for all behavioral inhibition task stages; and the natural language summary generated based on the above indicators; all raw EEG signals, ECG signals, skin conductance signals and eye movement data are discarded after feature extraction is completed on the local edge device, and only the desensitized statistical indicators and the training report are retained.

Images