Project Overview
Problem: Early signs of mental stress or mental illness often go undetected until they escalate.
AI Component: Sentiment analysis, anomaly detection, and predictive modeling analyze behavioral data, text inputs, and physiological signals.
Solution: A cloud-based AI system that correlates text emotions, sleep patterns, heart-rate variability, and activity patterns to detect early risk indicators.
Impact: Helps in early intervention and reduces global mental health crises.
Key Features
- Real-time sentiment analysis from typed text or voice transcripts
- Anomaly detection on wearable sensor signals
- Predictive mood forecasting (next 24–72 hours)
- Safety alerts and check-in prompts
- User dashboard + clinician view
No diagnosis — system assists professionals and individuals with insights.
Architecture
- Frontend: React app for journaling + mood dashboard
- Backend: Python (FastAPI) — ingestion, NLP scoring, signal processing
- NLP: TensorFlow sentiment and emotional-tone classifier
- Wearables: HRV, sleep, steps, stress metrics via device APIs
- Cloud: BigQuery / AWS Redshift for analytics + anomaly detection
Explainable AI modules highlight text patterns or physiological irregularities contributing to alerts.
Sample Dataset — User Profiles
| UserID | Age | Occupation | Region | Last Check-In |
|---|---|---|---|---|
| MH1001 | 19 | Student | Bangalore | 2025-11-15 |
| MH1002 | 27 | Software Engineer | Delhi | 2025-11-16 |
| MH1003 | 33 | Designer | Mumbai | 2025-11-14 |
| MH1004 | 41 | Teacher | Pune | 2025-11-12 |
Sample Dataset — Sentiment & Text Analysis
| UserID | Text Entry | Emotion | Sentiment Score | Risk Indicator |
|---|---|---|---|---|
| MH1001 | "Feeling overwhelmed with exams." | Stress | -0.62 | Medium |
| MH1002 | "Everything feels pointless this week." | Depressive | -0.80 | High |
| MH1003 | "Had a productive day, feeling good." | Positive | 0.70 | Low |
| MH1004 | "Not sleeping well lately." | Fatigue | -0.40 | Medium |
Sample Dataset — Wearable Sensor Metrics
| UserID | HRV (ms) | Resting HR | Sleep (hrs) | Steps | Anomaly |
|---|---|---|---|---|---|
| MH1001 | 65 | 78 | 5.1 | 3200 | Yes |
| MH1002 | 52 | 85 | 4.4 | 2100 | Yes |
| MH1003 | 90 | 70 | 7.2 | 8800 | No |
| MH1004 | 72 | 74 | 6.0 | 5100 | No |
Low HRV + poor sleep + negative sentiment often correlates with early stress patterns.
Sample Dataset — Predictive Risk Scores
| UserID | 24h Mood Forecast | Risk Level | Model Confidence |
|---|---|---|---|
| MH1001 | -0.58 | Medium | 0.84 |
| MH1002 | -0.82 | High | 0.91 |
| MH1003 | 0.65 | Low | 0.88 |
| MH1004 | -0.31 | Medium | 0.79 |
Forecast generated via LSTM TensorFlow model + wearable trend regression.
ML Logs & Explainability
| LogID | Model | Action | Time | Factors |
|---|---|---|---|---|
| LOG-801 | NLP-Sentiment-v3 | Analyze MH1002 text | 2025-11-16 | Keywords: "pointless", low energy patterns |
| LOG-802 | Wearable-Anomaly-v2 | HRV anomaly detection | 2025-11-15 | HRV=52ms, Sleep=4.4hrs |
| LOG-803 | LSTM-Mood-v4 | Forecast MH1001 mood | 2025-11-15 | Recent sentiment ↓, sleep ↓ |
| LOG-804 | FairnessCheck-v1 | Cohort bias scan | 2025-11-14 | No demographic imbalance detected |
Explainable AI helps clinicians understand why a risk alert was triggered.
Ethics & Safety Considerations
- Not a replacement for clinical diagnosis — only early-warning assistance
- Strict data privacy for text and wearable streams
- User control: delete data, download reports, disable monitoring
- Crisis-escalation protocol for high-risk signals
- Bias testing across age, gender, and region