Inspiration
The inspiration for SeizureSense came from a deeply personal experience. Witnessing one of our close friends experiencing the constant anxiety and uncertainty that comes with epilepsy, we asked ourselves: "How can technology provide peace of mind when every second matters?"
Traditional seizure detection often relies on physical observation or requires expensive, hospital-grade equipment. We envisioned a system that could:
- Detect seizures in real-time using accessible EEG technology
- Provide instant alerts to caregivers, even when they're not physically present
- Offer comprehensive monitoring through an intuitive dashboard
- Be affordable and accessible for home use
The intersection of biomedical signal processing, real-time machine learning, and IoT hardware presented an obvious technical challenge that could make a real difference in people's lives.
What it does
SeizureSense is a real-time EEG monitoring and seizure detection system that:
Monitors Brain Activity: Continuously processes EEG signals at 178 Hz, analyzing brainwave patterns in real-time
Detects Seizures: Uses an XGBoost machine learning model (98.74% accuracy) to predict seizure probability from extracted features
Sends Instant Alerts: When seizure probability exceeds 70%, the system:
- Sends WhatsApp alerts to caregivers with GPS location
- Provides specific details about the seizure and its nature
- Displays alerts on the dashboard
Provides Real-Time Visualization: Interactive React dashboard showing:
- Live seizure probability charts
- Heart rate monitoring
- Detection history and statistics
- System status and controls
Tracks Location: Automatically includes GPS coordinates in emergency alerts, helping caregivers locate patients quickly
Learns from Feedback: Local feedback system allows users to report false positives/negatives for continuous improvement
How we built it
Phase 1: Data Processing Pipeline We started by building a robust EEG signal processing pipeline:
- Signal acquisition - 178 Hz sampling rate, 1-second windows
- Feature extraction - 11 features total:
- Model inference - XGBoost classification in real-time
Phase 2: Backend Infrastructure The Flask backend serves multiple purposes:
- REST API for status, controls, and data retrieval
- WebSocket server for real-time data streaming
- ML inference engine processing EEG windows continuously
- Alert system integrating Twilio (WhatsApp) and ESP32 (buzzer)
- Location services with GPS and IP-based fallback
Phase 3: Frontend Dashboard The React frontend provides:
- Real-time charts using Chart.js showing seizure probability over time
- Live metrics including heart rate and detection status
- Control panel for starting/stopping monitoring
- Alert history and feedback system
- WebSocket integration for live updates without polling
Phase 4: Alert System We integrated multiple alert mechanisms:
- WhatsApp alerts via Twilio with location sharing and Google Maps links
- Dashboard notifications for immediate visual feedback
- Alert cooldowns to prevent spam
Phase 5: Location Services Added location sharing for emergency situations:
- Browser locater for GPS-level accuracy
- IP-based fallback for approximate location
- Google Maps integration for easy navigation to patient location
Challenges we ran into
- Real-Time Performance: Processing 178 samples per second while maintaining < 100ms latency for medical-grade responsiveness
Our Solution:
- Optimized feature extraction using NumPy vectorization
- Implemented data buffering and batch processing
- Used efficient data structures to minimize memory overhead
- False Positive Reduction: Initial model had high false positive rate, causing alert fatigue and reducing trust in the system
Our Solution:
- Implemented debouncing requiring 3 consecutive detections above threshold
- Added cooldown periods
- WebSocket Reliability: Maintaining stable connections during long monitoring sessions, especially with unreliable networks
Our Solution:
- Added connection status indicators on frontend
- Heartbeat/ping mechanisms to detect dead connections
- Hardware Integration: ESP32 communication failures, port detection issues across different operating systems, and serial port conflicts
Our Solution:
- Auto-detection of serial ports across macOS, Windows, and Linux
- Comprehensive error handling for hardware failures
- Clear user feedback when hardware is unavailable
- Instructions to close Arduino Serial Monitor before use
- Location Privacy: Balancing location accuracy with user privacy concerns
Our Solution:
- Location only sent during actual seizure events (not continuously streamed)
- Clear privacy documentation explaining data usage
- Option to use approximate IP-based location instead of GPS
- All location data stored temporarily in memory, not persisted
- Model Deployment: Ensuring consistent model performance across different environments and handling edge cases
Our Solution:
- Model versioning and validation on test datasets
- Comprehensive logging for debugging inference issues
- Fallback mechanisms when model fails to load
Accomplishments that we're proud of
98.74% Model Accuracy: Achieved high accuracy with XGBoost while maintaining real-time performance
Real-Time Processing: Built a system that processes 178 Hz EEG data in real-time with minimal latency, meeting medical monitoring requirements
Multi-Channel Alert System: Implemented three alert mechanisms (WhatsApp, ESP32 buzzer, dashboard) with intelligent cooldown management
Location Sharing: Added GPS location to emergency alerts, making it easier for caregivers to respond quickly
Robust Error Handling: System gracefully handles hardware failures, network issues, and missing dependencies without crashing
What we learned
- Biomedical Signal Processing
EEG signals are incredibly complex and noisy. We learned to extract meaningful features from raw brainwave data:
Time-domain features:
- Statistical measures: mean, standard deviation, RMS
- Higher-order moments: kurtosis, skewness
- Peak-to-peak amplitude analysis
- Real-Time Machine Learning
We discovered that model selection is critical for real-time applications. After testing Random Forest, Neural Networks, and XGBoost, we chose XGBoost for its perfect balance:
- 98.74% accuracy on test data
- < 50ms inference time per sample
- Low memory footprint for edge deployment
Key learnings:
- Debouncing is essential to reduce false positives
- Threshold tuning balances sensitivity and specificity
- Feature engineering matters more than model complexity for this use case
- Real-time constraints require different optimization strategies than batch processing
- Medical Technology Standards
We learned about critical requirements for medical devices:
- False positive minimization: Debouncing and threshold tuning
- System reliability: Graceful handling of hardware failures
- Data privacy: Local processing and secure communication
- User interface design: Clear, non-alarming visual feedback
What's next for SeizureSense
Short-Term Enhancements
- Mobile App: Native iOS/Android app for on-the-go monitoring
- Multi-Contact Alerts: Alert multiple caregivers simultaneously
- Historical Trends: Long-term seizure pattern analysis and visualization
- Seizure Diary: Manual logging with trigger correlation analysis
Advanced Features
- Multi-Model Ensemble: Combine XGBoost, LSTM, and CNN for improved accuracy
- Seizure Type Classification: Distinguish between tonic-clonic, absence, focal seizures
- Personalized Models: Fine-tune model per user using feedback data
Clinical Integration
- EHR Integration: Connect with Epic, Cerner for healthcare provider access
- Doctor Portal: Web interface for neurologists to review patient data
- Clinical Validation: Partner with medical institutions for validation studies
- Regulatory Compliance: Work toward FDA/CE marking for medical device approval
All-in-all, SeizureSense represents our journey from concept to working prototype, combining machine learning, biomedical engineering, and software development to create something that could genuinely help people live safer, more independent lives.