Dynamic Spatio-Temporal Graph Neural Network for Early Detection of Pornography Addiction in Adolescents Based on Electroencephalogram Signals

This study proposes a Dynamic Spatio-Temporal Graph Neural Network (DST-GNN) that integrates Phase Lag Index-based Graph Attention Networks and Bidirectional Gated Recurrent Units to achieve 71% F1-score accuracy in objectively detecting adolescent pornography addiction from EEG signals, significantly outperforming conventional machine learning baselines.

The Gist

Imagine your brain is a bustling city with 19 major communication towers (the EEG electrodes) sending signals back and forth. In a healthy city, the traffic flows smoothly, and the towers talk to each other in a balanced, organized way. But in a city struggling with addiction, the traffic patterns get chaotic, the towers start shouting at each other in specific, repetitive loops, and the "city planners" (the brain's control centers) lose their ability to regulate the flow.

This paper is about building a super-smart detective that can spot these chaotic traffic patterns in teenagers to detect pornography addiction early, without needing them to admit it or fill out a survey (which they might lie about due to shame).

Here is the breakdown of how they built this detective, using simple analogies:

1. The Problem: Why We Need a New Detective

Traditionally, doctors ask teenagers, "Do you watch too much porn?" This is like asking a thief if they stole a cookie. They might say "No" because they feel guilty or embarrassed. This is called self-report bias.

Previous computer programs tried to analyze brain waves (EEG) to find the answer, but they were like static snapshots. They took a picture of the brain's traffic at one single moment. But addiction isn't a still photo; it's a movie. The brain's connections change every millisecond when a person is exposed to triggers. Old computers missed these dynamic changes.

2. The Solution: The "Dynamic Spatio-Temporal" Detective

The researchers built a new AI called DST-GNN. Think of it as a high-tech surveillance system that does two things at once:

• The "Spatial" Part (The Map): It looks at where the signals are coming from. It uses a special tool called PLI (Phase Lag Index) to draw a map of which towers are talking to which. It's like a traffic controller who knows exactly which towers are connected and how strong that connection is, ignoring the static noise (like wind or rain) that usually messes up the signal.
• The "Temporal" Part (The Movie): It watches how those connections change over time. It uses a memory unit called BiGRU (Bidirectional Gated Recurrent Unit) to remember the past and predict the future of the signal. It's like a detective who doesn't just look at a traffic jam but watches the whole rush hour to see if the pattern is normal or suspicious.

3. The Training: Teaching the Detective

They taught this AI using data from 14 teenagers (7 with addiction, 7 healthy).

• The Test: The teens watched different things: happy videos, sad videos, scary videos, and even pornographic images (the "Executive Task").
• The Challenge: The AI had to guess who was addicted just by looking at their brain waves, without being told who was who.
• The Result: The AI got it right 71% of the time (which is a huge jump from the old methods that barely got 35% right). Most importantly, it caught 85% of the addicted teens. In medical screening, catching the "sick" people is more important than avoiding false alarms, so this is a massive win.

4. The "Aha!" Moments: What Did the AI Learn?

The researchers didn't just want a black box; they wanted to know why the AI made its decisions. They asked the AI to explain its logic, and here is what it found:

• The "Frontal-Central" District: The AI noticed that the addiction mostly showed up in the front and center of the brain (areas Fz, Cz, C3, C4). Think of this as the brain's "CEO office" responsible for impulse control. In addicted teens, this office is overworked and confused.
• The "Beta" Frequency: The AI found that the brain waves in the Beta range (fast, busy waves) were the biggest clue. It's like the brain is running a high-speed race when it shouldn't be.
• The "Secret Code" (Cz-T7): The most exciting discovery was a specific connection between two points (Cz and T7) that looked different in addicted teens regardless of what they were watching. Even when they were just sitting with their eyes closed, this connection was "broken." This is like finding a unique fingerprint that proves someone is addicted, even if they aren't currently looking at porn. This means doctors might eventually be able to screen kids without showing them any triggering images at all.

5. Why This Matters

This research is like giving parents and doctors a new pair of glasses that can see the invisible signs of addiction in a teenager's brain.

• No Shame: It removes the need for teenagers to confess, reducing stigma.
• Early Intervention: It can catch the problem before it ruins their development.
• Better Treatment: By knowing exactly which brain parts are struggling, doctors can design better therapies (like neurofeedback) to "retrain" the brain's traffic patterns.

In a nutshell: The researchers built a smart, time-aware AI that reads brain waves like a movie instead of a photo. It found that addicted teenagers have a specific, chaotic "traffic jam" in their brain's control center, and it can spot this pattern with high accuracy, offering a hopeful new way to protect children from addiction.

Technical Summary

1. Problem Statement

Adolescent pornography addiction is a growing public health concern in Indonesia, with high exposure rates (66.6% of boys and 62.3% of girls aged 13–17). Current screening methods, such as the Young's Pornography Addiction Screening Test (YPAST), rely on self-reports, which are prone to social desirability bias and subjective guilt. Furthermore, conventional machine learning approaches (e.g., SVM, Random Forest) typically utilize static features, failing to capture the dynamic functional connectivity of the brain that fluctuates temporally during exposure to addictive stimuli. There is a critical need for an objective, neurobiological biomarker-based detection system that models both spatial brain connectivity and temporal dynamics.

2. Methodology

Dataset and Preprocessing

• Data Source: A public dataset from Mendeley Data (Kang et al., 2021) containing EEG recordings from 14 adolescents (7 addicted, 7 healthy controls), aged 13–15.
• Acquisition: 19-channel EEG (10-20 system) sampled at 250 Hz.
• Experimental Conditions: 9 tasks including Eyes Closed/Open, emotional states (Happy, Calm, Sad, Fear), cognitive tasks (Memorize, Recall), and an Executive Task (ET) involving pornographic stimuli.
• Preprocessing:
  • Bandpass filtering (0.5–45 Hz) using a 4th-order Butterworth filter.
  • Notch filtering at 50 Hz to remove powerline interference.
  • Z-score normalization per channel.
  • Windowing into 30 temporal windows.

Feature Extraction

Two categories of features were extracted per channel (19 channels × 9 features = 171 node features):

1. Power Spectral Density (PSD): Calculated via Welch's method across 5 bands: Delta, Theta, Alpha, Beta, and Gamma.
2. Hjorth Parameters: Activity, Mobility, Complexity, and Mean Amplitude.

Graph Construction

• Metric: Phase Lag Index (PLI) was used to measure functional connectivity. PLI was chosen for its robustness against volume conduction artifacts compared to standard coherence metrics.
• Graph Topology: A graph $G=(V, E)$ was constructed where nodes ($V$) are the 19 EEG channels and edges ($E$) are weighted by PLI values.
• Thresholding: The top 50% of strongest connections were retained to form the adjacency matrix.

Model Architecture: DST-GNN

The proposed Dynamic Spatio-Temporal Graph Neural Network consists of three integrated components:

1. Spatial Encoder (GAT): A 2-layer Graph Attention Network (GAT) with 2 attention heads and hidden dimension 64. It models spatial dependencies between EEG channels using adaptive attention weights to capture functional connectivity.
2. Temporal Encoder (BiGRU): A 2-layer Bidirectional Gated Recurrent Unit (BiGRU) with hidden dimension 128. It captures the temporal dynamics of the graph sequences over time.
3. Classifier: A Multi-Layer Perceptron (MLP) with dropout (0.5) for binary classification (Addicted vs. Healthy).

Evaluation Strategy

• Validation: Leave-One-Subject-Out Cross-Validation (LOSO-CV) with 3 random seeds (42, 123, 456) to ensure generalizability across subjects.
• Baselines: Compared against Logistic Regression, SVM (RBF), Random Forest, XGBoost, and MLP.
• Ablation Study: Tested the impact of removing the temporal component (BiGRU) and replacing the PLI graph with a fully connected graph.

3. Key Contributions

1. Novel Architecture: First application of a Dynamic Spatio-Temporal GNN (integrating GAT and BiGRU) specifically for pornography addiction detection in adolescents.
2. Robust Graph Construction: Integration of Phase Lag Index (PLI) to construct the brain connectivity graph, effectively mitigating volume conduction artifacts common in EEG analysis.
3. Objective Biomarkers: Identification of specific neurobiological markers (brain regions and connectivity patterns) that serve as objective screening tools, reducing reliance on subjective self-reporting.
4. Trait-Level Discovery: Discovery of a condition-independent biomarker (Cz–T7 connectivity) that could enable detection without direct exposure to pornographic stimuli.

4. Results

Performance Metrics

• DST-GNN Performance: Achieved an average F1-Score of 71.00% ± 12.10% and a Recall of 85.71%.
• Improvement: Represents a 104% improvement in F1-Score compared to the best baseline (Logistic Regression: 34.73%).
• Recall Significance: The high recall (85.71%) is critical for medical screening to minimize false negatives (missing addicted individuals).
• Baseline Comparison: All conventional baselines struggled with recall (<30%), confirming the necessity of modeling temporal dynamics and connectivity.

Ablation Study

• Temporal Dynamics: Removing the BiGRU component decreased the F1-Score by 21%, proving the importance of capturing time-varying brain states.
• Graph Construction: Replacing the PLI-based graph with a fully connected graph caused a drastic 57% drop in F1-Score, validating that functional connectivity structure is essential.

Explainability & Biomarkers

• Dominant Regions: Frontal and central regions (Fz, Cz, C3, C4) were identified as the most discriminative.
• Feature Importance:
  • Beta Band (13–30 Hz): Contributed 58.9% (associated with arousal and cognitive control).
  • Hjorth Parameters: Contributed 31.2% (specifically complexity and activity).
• Trait-Level Biomarker: The Cz–T7 connectivity showed consistent differences across multiple experimental conditions (Eyes Closed, Eyes Open, Happy, Memorize), suggesting it is a stable trait marker for addiction rather than a state-dependent response.
• Connectivity Patterns: Addicted subjects showed decreased global connectivity but increased local clustering in right frontal-temporal regions, consistent with allostatic theory of addiction.

5. Significance and Implications

• Child Protection: Provides an objective, neuroscience-based screening tool to support Law No. 35 of 2014 regarding Child Protection in Indonesia, allowing for early intervention.
• Clinical Utility: The high recall rate ensures that at-risk adolescents are not missed during screening. The identification of trait-level biomarkers (Cz-T7) suggests future screening could be performed without exposing minors to explicit content.
• Intervention Design: The localization of biomarkers to frontal-central regions can guide the development of targeted neurofeedback interventions to improve impulse control and emotional regulation.

6. Limitations and Future Work

• Sample Size: The study used a small dataset (N=14), which limits generalizability and contributes to high variance in results.
• Demographics: Limited to a specific age range (13–15) and a single cultural context (Indonesia).
• Real-time Application: The model has not yet been validated on real-time streaming data.
• Future Directions: Validation on larger, diverse datasets, development of real-time detection systems, and integration with neurofeedback protocols.