SafeSpace AI

Multimodal, Explainable, Real-Time Stress Detection Platform

• Overview
• System Architecture
• Data Modalities
• Sensing Hardware
• AI Models & Fusion
• Explainable AI (XAI)
• Results & Evaluation
• How to Run
• Downloads & Outputs
• Future Work
• Contributors

Overview

SafeSpace AI is an end-to-end, cloud-ready framework for real-time, explainable, multimodal stress detection.
The system fuses physiological biosignals, voice cues, and self-reported questionnaire data via advanced AI, and delivers actionable, plain-language insights for both users and clinicians.

• Live Demo: safespaceai.vercel.app
• Wearable hardware, transparent AI, instant results.

System Architecture

flowchart LR
    A[ESP32 Wearable: ECG, EDA, EMG, Temp] --> B[Physio Model]
    C[Web/Mic: Voice Input] --> D[Voice Model]
    E[Web: DASS-21 Survey] --> F[Survey Model]
    B --> G[Late Fusion + XAI]
    D --> G
    F --> G
    G --> H[Web Dashboard & API Output]

Loading

Components:

• ESP32-based wearable hardware
• FastAPI cloud backend (Python)
• Next.js/React web dashboard
• Modular ML stack for each modality

Data Modalities

Modality	Features (examples)	Data Source
Physiological	ECG, EDA, EMG, Body Temp, HRV, freq. feats	ESP32 Wearable
Voice	MFCCs, pitch, energy, prosody, chroma	Web/microphone
Self-Assessment	DASS-21 Stress Subscale (7 items, web)	Web form

Sensing Hardware

• ESP32 microcontroller with wireless data streaming
• Sensors: ECG, EDA, EMG, temperature
• Rechargeable battery (6+ hours use)
• Wearable form factor: ergonomic patch/armband
• Plug-and-play with SafeSpace AI backend

AI Models & Fusion

• Biosignals: Random Forest, SVM, XGBoost, Extra Trees (WESAD data)
• Voice: CNN-BiGRU-Attention (IEMOCAP → RAVDESS)
• Survey: Stacking ensemble (RF, LR, NB + meta LR)
• Late Fusion:
  • Weighted aggregation by model confidence and input quality
  • Typical weights: 60% biosignals, 25% survey, 15% voice
  • Robust to missing/partial data

Explainable AI (XAI)

• SHAP: Feature attributions for biosignals & survey
• LIME: Explains voice/tabular predictions
• Attention Visualization: Neural model insight (voice)
• LLM Integration: Plain-language, actionable summaries for each prediction
• Full XAI feedback in user dashboard

Results & Evaluation

Modality	Best Accuracy	Macro AUC	Datasets
Physiological	88.3%	0.97	WESAD
Voice	87.8%	0.97	IEMOCAP, RAVDESS
DASS-21 Survey	85.0%	0.96	DASS-21
Fusion	83.9%	0.94	(All Combined)

• Fusion improves detection of “medium” stress and overall reliability.
• All outputs include SHAP/LIME/Attention plots and natural-language explanations.

How to Run

1. Hardware & Data

• Flash ESP32, attach sensors as per /hardware_setup.md.
• Connect to web dashboard (/frontend), pair device & mic.

2. Backend

pip install -r requirements.txt
uvicorn backend.main:app --reload

3. Frontend

cd frontend
npm install
npm run dev

• Visit localhost:3000 or safespaceai.vercel.app

4. Model Training

• Run scripts in /models (e.g., python train_physio.py)
• Pretrained weights included for demo.

Downloads & Outputs

• Dashboard: Real-time predictions, modality contributions, XAI explanations
• Exports:
  • Class probabilities CSV
  • Confusion matrix & ROC (PNG)
  • SHAP/LIME plots
  • Per-session summary reports

Future Work

• Personalized/online learning per user
• Larger, more diverse datasets; multi-lingual support
• Federated and on-device privacy-preserving inference
• Full clinical & occupational trials

Contributors

• Arsh Chatrath
• Devansh Wadhwani
• Durvish Khurana
• Kabir Oberoi
• Samya Agarwal
• Khushpreet Kaur
• Lakshita

SafeSpace AI bridges advanced hardware, modern ML, and explainable, real-time analytics for trustworthy mental health support.

Live at safespaceai.vercel.app. For documentation and details, see /docs.