Objective mood-episode risk from your Apple Watch — clinically informed, privacy-first, open-source.

The clinical problem: No objective tool exists for predicting mood episodes or distinguishing unipolar from bipolar depression or borderline personality disorder. Clinicians rely on subjective recall; patients often seek help after crises begin.

Big Mood Detector analyzes your Apple Health data to predict mood episode risk using AI. Two complementary models:

• PAT transformer screens for current depression.
• XGBoost predicts tomorrow's depression/mania/hypomania risk.

When used together (--ensemble), see how your mood state evolves from today to tomorrow.

Built by a clinical psychiatrist, implementing published research. Runs 100% on-device, no cloud, no signup.

Current status: Research prototype — the first of its kind, but not yet clinically validated.

Our breakthrough:

• Early detection: Spot mood episode risk before symptoms spiral
• Two applications: Current depression screening (PAT) + next-day episode prediction (XGBoost)
• Objective data: Complement clinical assessment with continuous behavioral biomarkers
• Research foundation: First implementation combining two breakthrough papers:
  • PAT: Dartmouth Foundation Model (arXiv)
  • XGBoost: Nature Digital Medicine 2024 (GitHub)
• Privacy-first: Runs entirely on your device — your data never leaves your machine

For researchers: Validate these approaches across populations, build the evidence base for digital mental health. See PAT model training details for replication.

Consistent environment with all dependencies

# 1. Create security credentials
cat > .env << EOF
SECRET_KEY=$(python3 -c 'import secrets; print(secrets.token_urlsafe(32))')
API_KEY_SALT=$(python3 -c 'import secrets; print(secrets.token_urlsafe(32))')
EOF

# 2. Start services
docker-compose up -d api redis

# 3. Export Apple Health data (Settings → Health → Export)
#    Place export.xml in: data/input/apple_export/

# 4. Analyze your data
docker run --rm \
  -e BIGMOOD_DATA_DIR=/app/data \
  -v "$(pwd)/data:/app/data" \
  -v "$(pwd)/model_weights:/app/model_weights:ro" \
  big-mood-detector:latest \
  predict /app/data/input/apple_export/export.xml --report

Requires Python 3.12+

# 1. Install (numpy first to avoid conflicts)
pip install 'numpy<2.0'
pip install big-mood-detector

# 2. Export Apple Health data (Settings → Health → Export)
#    Place export.xml in: data/input/apple_export/

# 3. For large files (>100MB), scan first to check data availability
big-mood predict data/input/apple_export/export.xml --scan

# 4. Analyze your data (research purposes)
big-mood predict data/input/apple_export/export.xml --report

# 5. With temporal analysis (NOW vs TOMORROW)
big-mood predict data/input/apple_export/export.xml --ensemble --report

# 6. Auto-window selection (enabled by default in v0.5.6+)
# Automatically finds optimal data windows for both models
big-mood predict data/input/apple_export/export.xml --report

# 7. To disable auto-window and specify dates manually
big-mood predict data/input/apple_export/export.xml --no-auto-window --date-range 2024-01-01:2024-03-31

See full output in data/output/clinical_report.txt

Need model weights? Please contact me for a convenient Google Drive link containing:

• Pre-trained PAT transformer weights
• Fine-tuned depression detection head
• XGBoost model files
• Sample data for testing

This ensures you can run the application immediately without training models from scratch.

Your Apple Health Data
      ↓
┌─────────────────────┐
│   Past 7 Days       │ ← PAT (transformer) analyzes activity patterns
├─────────────────────┤
│   Past 30 Days      │ ← XGBoost models circadian rhythms  
└─────────────────────┘
      ↓
Research Risk Scores (Not Diagnostic)

PAT = transformer AI, XGBoost = gradient boosting, ensemble = enhanced reliability.

New! Fast data scanning to preview what's available before processing:

# Scan a large export (takes seconds, not minutes)
big-mood predict export.xml --scan

# Example output:
📊 Data Summary:
• Total records: 8,755,251
• Heart Rate: 5,074,424 records
• Step Count: 3,513,756 records
• Sleep Analysis: 3,608 records

✅ Available Features:
• depression_risk: Depression risk prediction (XGBoost)
• mania_risk: Mania/hypomania risk prediction (XGBoost)

⚠️ Unavailable Features:
• hrv_analysis: Missing required type: HKQuantityTypeIdentifierHeartRateVariabilitySDNN

Files over 100MB will automatically prompt to scan first. This prevents waiting 10+ minutes only to discover missing data.

Processing Speed:

• 365 days of data: 17 seconds
• Memory usage: <100MB for any file size
• Parsing throughput: 33MB/s

Model Accuracy (from original research):

• Current depression screening (PAT, general population):
  • Depression detection: 0.593 AUC (US NHANES 2013-14)¹
• Next-day episode prediction (XGBoost, mood disorder patients):
  • Depression episodes: 0.80 AUC (Korean cohort, MDD+BD patients)²
  • Manic episodes: 0.98 AUC (Korean cohort, BD patients)²
  • Hypomanic episodes: 0.95 AUC (Korean cohort, BD patients)²

¹Ruan et al., 2024 | ²Lim et al., 2024

1. Clinical innovation: First open-source tool combining two state-of-the-art approaches to predict mood episodes from wearable data

2. Scientific rigor: Faithful implementation of published algorithms:

   • XGBoost circadian model from Seoul National University (Nature Digital Medicine 2024)
   • PAT transformer from Dartmouth (first foundation model for actigraphy)

3. Privacy breakthrough: No cloud dependency, no data collection — your mental health data stays private

4. Intelligent data handling: Auto-window selection (v0.5.6+) automatically finds optimal analysis periods:

   • Handles sparse Apple Watch data (gaps are OK for XGBoost)
   • Prioritizes overlapping windows where both models can run
   • Clear feedback on data availability and model selection

5. Open research: Complete transparency enables validation, improvement, and trust

Population specificity:

• XGBoost: Trained on 168 Korean adults (18-35y) with mood disorders
• PAT: Pre-trained on 21,538 US adults, fine-tuned on 2,800 with PHQ-9 scores

Performance constraints:

• Current depression screening: Moderate accuracy (0.593 AUC)
• Next-day episode prediction: High accuracy but limited to Korean cohort (0.80-0.98 AUC)
• No validation across ethnicities, age groups, or comorbid conditions
• Research tool only — not FDA approved or clinically validated

Critical research needs:

• 🏥 Clinical validation across diverse populations
• 🌍 Multi-ethnic, multi-age cohort studies
• 📱 Integration with additional wearable devices
• 🧪 Improving transformer model accuracy

See CONTRIBUTING.md for guidelines.

Apache 2.0 - Use freely, even commercially. See LICENSE.

Built on pioneering research:

• XGBoost models: Seoul National University, Korea University, KAIST (Lim et al., 2024)
• PAT foundation model: Dartmouth College (Ruan et al., 2024)

Have feedback? Open an issue or see discussions.