Research & educational use only
Risk assessment
Enter patient or subject data to compute an AI-powered addiction risk score
⚠️ This tool is for research and educational purposes only. Not validated for clinical diagnosis. AI reasoning is grounded in published COGA / NIAAA neuroscience — not trained on restricted dbGaP EEG recordings.
Subject profile
Cognitive performance
From neuropsychological battery or behavioral task results
Reaction time (ms)Norm: 250–380 ms
320 ms
Task accuracy (%)Norm: 85–95%
88%
Commission errors (%)Norm: <12%
10%
Inhibitory control scoreHigher = better. Norm: 70–90
80
Working memory scoreHigher = better. Norm: 75–90
82
Self-report / clinical scales
Impulsivity (BIS-11 style)Higher = more impulsive. Norm: 30–45
38
Reward sensitivityHigher = more sensitive. Norm: 45–60
52
AUDIT score0–7 low · 8–15 hazardous · 16+ harmful
4
About this system
Architecture, evidence base, and intended use
Evidence base

This tool is grounded in findings from the Collaborative Study on the Genetics of Alcoholism (COGA) — a multi-site longitudinal study funded by NIAAA spanning 30+ years. Key biomarkers used:

  • P300 amplitude — reduced in individuals with AUD and their at-risk offspring before first drink (Begleiter et al. 1984; Jones et al. 2006)
  • Beta power — elevated frontal beta linked to GABRA2 on chromosome 4, associated with AUD (Rangaswamy et al. 2002, 2004; Porjesz et al. 2002)
  • Theta coherence — fronto-parietal theta synchrony reduced in AUD; CHRM2 on chromosome 7 implicated (Porjesz & Rangaswamy 2007)
  • Delta/theta power — weaker during oddball task in AUD and high-risk relatives (Jones et al. 2006)
  • Machine learning — EEG measures used to predict AUD development (Kinreich et al. 2021)
COGA / cogastudy.org NIAAA / niaaa.nih.gov dbGaP phs000125 Begleiter 1984 Porjesz 2002 Kinreich 2021
System architecture
1
Data layer — data_loader.py
Simulates COGA-style EEG epochs (19 channels, 256 Hz), behavioral scores, and clinical labels. Replace simulate_coga_dataset() with your real dbGaP download.
2
Preprocessing — preprocessing.py
60 Hz notch filter → 1–45 Hz bandpass → amplitude-based artifact removal → per-channel z-score normalization → KNN imputation for missing behavioral values.
3
Feature extraction — feature_extraction.py
217 features: spectral power (5 bands × 19 channels × absolute + relative), P300/N100 ERP components, fronto-parietal coherence across 5 channel pairs × 3 bands, and 7 behavioral scores.
4
Multimodal model — model.py
Stacked ensemble: EEG branch (GradientBoosting) + Behavioral branch (GradientBoosting) → RandomForest meta-learner for binary classification + GradientBoostingRegressor for severity score (0–10).
5
Training — train.py
70/15/15 train/val/test split · 5-fold stratified cross-validation · data augmentation (noise injection) · model serialized to model.pkl.
6
Evaluation — evaluate.py
ROC curve, confusion matrix, CV band plot, severity scatter, attention-weight heatmap (channel × band), feature importance bars. All saved to outputs/dashboard.png.
Clinical applications
  • Early risk detection — P300 attenuation is detectable in adolescents before first drink; enables preventive intervention in high-risk families
  • Disorder differentiation — extend label set to AUD / ADHD / MDD / Healthy for multiclass classification of overlapping cognitive phenotypes
  • Treatment monitoring — run severity regression longitudinally to track whether a patient's neural profile is converging toward the healthy distribution during intervention
  • Genetic risk stratification — integrate polygenic risk scores (GABRA2, CHRM2 variants) as additional features
Running the Python pipeline
# Install dependencies
pip install numpy scipy scikit-learn matplotlib seaborn mne pandas

# Run full pipeline (data → train → evaluate → dashboard)
cd neuro_addiction_ai
python main.py

# Outputs saved to neuro_addiction_ai/outputs/
# model.pkl — serialized trained model
# results.json — all metrics and feature importances
# dashboard.png — full evaluation visualization
Important limitations
  • The web app AI reasoning uses Claude's knowledge of COGA/NIAAA literature — it is not a model trained on the actual COGA EEG recordings (those require NIH dbGaP data access approval)
  • The Python pipeline trains on simulated data that replicates COGA statistical signatures — performance on real data will differ
  • Perfect cross-validation scores (AUC=1.0) are expected on clean simulated data and are not representative of real-world performance
  • This tool must not be used for clinical diagnosis or treatment decisions