Gamming and Mental Health - Dataquest 26

A hackathon project that uses machine learning to analyze the relationship between gaming habits and mental wellbeing. Users fill out a short assessment about their gaming behavior and lifestyle; the system returns a personalized risk analysis across five wellbeing dimensions with explainable predictions.

---

What It Does

• Predicts risk levels for five mental health outcomes based on a user's gaming profile:
  • Sleep problems
  • Productivity issues
  • Social isolation
  • Dysregulation
  • Emotional problems
• Shows which specific habits (hours played, genre, spending, exercise, etc.) are driving each risk
• Presents results in a retro arcade-themed web interface with charts and scenario analysis

---

Project Structure

Dataquest26/
├── data/                          # Source dataset (1,000 records, 27 features)
├── mood/                          # Mood state classifier (Random Forest)
│   ├── train.py
│   ├── train_notebook.ipynb
│   └── model.joblib / model.pkl
├── train/                         # Multi-issue wellbeing models
│   ├── train_causal_grouped_model.py
│   ├── infer_grouped_model.py
│   ├── gamer_personas_kmeans.ipynb
│   ├── causal_detection_wellbeing*.ipynb
│   └── artifacts/                 # Trained model + metrics
├── analysis/                      # Causal pathway exploration
│   └── path_analysis.ipynb
└── frontend/                      # Next.js web application
    └── app/
        ├── page.tsx               # Landing page
        ├── assess/page.tsx        # User intake form
        ├── results/page.tsx       # Results + visualizations
        └── api/analyze/route.ts   # Inference API (calls Python)

---

Tech Stack

Layer	Technology
ML Models	scikit-learn (RandomForest, GradientBoosting, LogisticRegression)
Data	pandas, numpy
Notebooks	Jupyter
Frontend	Next.js 16, React 19, TypeScript
Styling	Tailwind CSS 4
Charts	Recharts

---

Dataset

data/Gaming and Mental Health.csv — 1,000 rows covering:

• Demographics (age, gender, years gaming)
• Gaming habits (daily hours, spending, genre, platform)
• Physical health (sleep hours, exercise, eye strain)
• Mental health outcomes (mood state, social isolation, addiction risk level)

---

Machine Learning Models (train/)

There are two scripts in the train/ folder that together form the core ML pipeline.

train_causal_grouped_model.py — Model Training

Trains a bundle of classifiers that predict five distinct wellbeing issues from a user's gaming and lifestyle inputs. The training pipeline works as follows:

1. Feature taxonomy — Input features are split into three groups:

   • Actionable causes: daily_gaming_hours, monthly_game_spending_usd, exercise_hours_weekly, game_genre, primary_game, gaming_platform
   • Context confounders: age, gender, years_gaming
   • Mediators (used in augmented track only): sleep, mood, social, and productivity metrics

2. Target definitions — Five binary classification targets are derived from the dataset:

   • sleep_problem: poor sleep quality or frequent sleep disruption
   • productivity_problem: poor academic/work performance or low productivity score
   • social_isolation_problem: high isolation score or low face-to-face social hours
   • emotional_problem: mood state is Depressed, Anxious, or Withdrawn
   • dysregulation_problem: frequent mood swings or continued gaming despite problems

3. Leakage prevention — For each issue, the features that directly define the target (e.g. sleep_quality for sleep_problem) are excluded from that model's input, preventing circular predictions.

4. Model selection — Three algorithms are trained and evaluated for each issue; the best by F1 + ROC-AUC is kept:

   • RandomForestClassifier (500 trees, balanced class weights)
   • GradientBoostingClassifier
   • LogisticRegression (balanced class weights)

5. Overall model — A separate top-level classifier (also best-of-three) predicts a composite wellbeing_problem flag (two or more issues present simultaneously).

6. Preprocessing pipeline — Numeric features are median-imputed and standard-scaled; categorical features are mode-imputed and one-hot encoded, all within a ColumnTransformer pipeline to prevent data leakage.

Output: train/artifacts/grouped_wellbeing_model.pkl — a pickle bundle containing all five issue models, the overall model, feature sets per issue, and the feature taxonomy.

Issue	Best Model	F1	ROC-AUC	Reliability
Sleep problem	GradientBoosting	0.809	0.672	Reliable
Social isolation	RandomForest	0.794	0.927	Reliable
Productivity problem	RandomForest	0.646	0.716	Reliable
Dysregulation	GradientBoosting	0.550	0.626	Weak
Emotional problem	LogisticRegression	0.374	0.559	Weak
Overall wellbeing	RandomForest	0.667	0.898	—

---

infer_grouped_model.py — Inference Engine

A command-line inference script that loads the trained model bundle and runs predictions for a single user profile. It is called by the Next.js API route at runtime.

What it does:

• Accepts a JSON payload (via --payload or --payload-file) containing the 9 required input fields
• Validates and range-clips all numeric inputs before prediction
• Runs the overall wellbeing model and all five issue-specific models
• Computes local feature contributions per issue using perturbation analysis: each feature is individually replaced with a neutral baseline value and the change in predicted probability is measured, giving an explainable breakdown of what is driving each risk
• Groups contributions into cause categories (gaming_load, gaming_spend, health_habits, game_context, context)
• Runs intervention scenarios on the overall risk model to show how specific behavioral changes (e.g. reduce gaming by 2h/day, add 2h exercise/week) would shift the predicted risk
• Outputs a single JSON object with overall risk, per-issue risks, top contributors, and scenario comparisons

Usage:

python train/infer_grouped_model.py \
  --artifact train/artifacts/grouped_wellbeing_model.pkl \
  --payload '{"age": 22, "daily_gaming_hours": 6, ...}'

---

Model Performance

Overall accuracy: 79.6%

Issue	F1 Score
Sleep problems	81.5%
Social isolation	79.4%
Productivity problems	~78%
Dysregulation	~78%
Emotional problems	~77%

Models were trained without pretrained neural networks or external APIs — all classical ML on the provided dataset.

---

Getting Started

Python / Model Training

pip install -r requirements.txt

# Train the wellbeing models
python train/train_causal_grouped_model.py

# Train the mood classifier
python mood/train.py

Web Application

cd frontend
npm install
npm run dev

Open http://localhost:3000 in your browser.

The app's /api/analyze route spawns a Python subprocess to run inference. Make sure Python dependencies are installed and the model artifacts exist in train/artifacts/ before starting the frontend.

---

How It Works

1. User fills out a 9-field form (age, gaming hours, spending, genre, platform, exercise, etc.)
2. The Next.js API route calls infer_grouped_model.py with the user's inputs
3. The model outputs risk levels and per-feature contribution scores for each wellbeing dimension
4. Results are displayed as pie charts, bar charts, and a scenario analysis breakdown

---

Notebooks

Notebook	Purpose
train/causal_detection_wellbeing.executed.ipynb	Core causal analysis and model training
train/causal_detection_wellbeing.localcontributors.executed.ipynb	Feature contribution analysis
train/causal_detection_wellbeing.personalized.executed.ipynb	Per-user prediction walkthrough
train/gamer_personas_kmeans.ipynb	K-means clustering of gamer archetypes
mood/train_notebook.ipynb	Mood state classifier training
analysis/path_analysis.ipynb	Causal pathway exploration