EpiWatch - Epileptic Motion Detection System

A motion-based epilepsy detection prototype designed to support student safety and real-time monitoring in educational environments. The system uses an ESP32-DOIT microcontroller and MPU-6050 6-axis accelerometer/gyroscope to capture motion data. Using Edge Impulse, the device runs a lightweight machine learning model directly on the ESP32 for real-time classification of seizure-like movement patterns.

⚠️ Important: This is a defensive safety tool designed to assist with monitoring and early alert systems. It's not intended as a medical diagnostic device but rather as an assistive technology for educational environments.

🎯 Project Overview

EpiWatch combines affordable hardware with advanced machine learning to create a wearable device that can detect abnormal motion patterns associated with seizure activity. The system provides immediate alerts through multiple channels while maintaining privacy through on-device processing.

Key Features

• Real-time Motion Analysis: 6-axis sensor data processed at 43Hz
• On-Device ML Inference: Edge Impulse model runs locally on ESP32
• Multi-Modal Alerts: buzzer and configurable output pins
• Low Power Design: Optimized for wearable applications
• Confidence Averaging: Reduces false positives through temporal smoothing
• Open Source: Complete hardware design and software stack

🔧 Hardware Components

Core Components

• ESP32-DOIT DevKit V1: Main microcontroller with WiFi/Bluetooth
• MPU-6050: 6-axis accelerometer/gyroscope sensor
• Alert Components: Buzzer (Pin 23)
• Power: 18650 lithium battery or USB power

Wiring Diagram

ESP32          MPU-6050
-----          --------
3.3V     →     VCC
GND      →     GND
GPIO 21  →     SDA
GPIO 22  →     SCL

ESP32          Alerts
-----          ------
GPIO 23  →     Buzzer

🏗️ 3D Printed Enclosure

The project includes a custom 3D-printable enclosure designed in FreeCAD:

• Files: CAD/Cover-Body.stl, CAD/Top_lid.FCStd
• Material: PLA plastic recommended
• Design: Compact wearable form factor with sensor access
• Assembly: Snap-fit design for easy assembly

📊 Machine Learning Pipeline

Data Collection

The system collects sensor data at 43Hz sampling rate:

• Accelerometer: X, Y, Z axes (±8G range)
• Gyroscope: X, Y, Z axes (±500°/s range)
• Window Size: 2 seconds (86 samples)
• Classes: Normal activity vs. seizure-like motion

Model Training

• Platform: Edge Impulse
• Processing: Spectral analysis for frequency domain features
• Architecture: Neural network with 20-neuron hidden layer
• Performance: >85% accuracy with <50ms inference time

🚀 Getting Started

Prerequisites

# Install PlatformIO
pip install platformio

# Install Python dependencies for data collection
pip install pandas numpy pyserial

# Install Edge Impulse CLI (optional)
npm install -g edge-impulse-cli

Building and Flashing

1. Clone the repository

   git clone <repository-url>
   cd epiwatch_deployment

2. Build the firmware

   pio run

3. Flash to ESP32

   pio run --target upload

4. Monitor output

   pio device monitor

📈 Data Collection & Training

Collecting Training Data

1. Normal Activity Data

   cd data-collector-tool
   python3 script_collector.py -p /dev/ttyUSB0 -l "normal" -o ../dataset/normal.csv -d 60

2. Seizure Simulation Data

   python3 script_collector.py -p /dev/ttyUSB0 -l "seizure" -o ../dataset/seizure.csv -d 60

Preparing Data for Edge Impulse

cd data
python3 prepare_for_edge_impulse.py -i ../dataset -o edge_impulse_data

Training the Model

Follow the comprehensive guide in data/edge_impulse_training_guide.md for detailed steps on:

• Creating an Edge Impulse project
• Uploading training data
• Configuring the impulse
• Training and optimizing the model
• Deploying to ESP32

🔍 System Operation

Real-time Detection

The device continuously monitors motion and provides:

• Confidence Scoring: Probability values for each prediction
• Temporal Averaging: Smooths predictions over 5 samples
• Threshold-based Alerts: Configurable seizure detection threshold (default: 0.7)
• Alert States: Visual and auditory feedback

Serial Output Example

Predictions (DSP: 3 ms, NN: 12 ms): normal: 0.95 seizure: 0.05
Average confidence over 5 samples: 0.04
✅ Normal activity detected.

📁 Project Structure

epiwatch_deployment/
├── src/                    # ESP32 source code
│   ├── main.cpp           # Main inference and alert logic
│   └── data_collection.cpp # Data collection mode
├── data-collector-tool/    # Python data collection scripts
├── data/                  # Edge Impulse training utilities
├── dataset/               # Raw training data (CSV)
├── CAD/                   # 3D printable enclosure files
├── sample/                # Hardware photos and documentation
└── lib/                   # Edge Impulse library

⚙️ Configuration

Adjustable Parameters

• Seizure Threshold: Confidence level for triggering alerts
• Sample Rate: Data collection frequency (default: 43Hz)
• Confidence Samples: Number of samples for temporal averaging
• Alert Pins: GPIO assignments of buzzer

Customization

The system can be extended with:

• WiFi/Bluetooth connectivity for remote alerts
• Data logging to SD card
• Integration with smartphone apps
• Multiple sensor fusion (heart rate, temperature)

🤝 Contributing

This is an open-source defensive safety project. Contributions are welcome for:

• Improved ML models and feature extraction
• Hardware optimizations and new sensor integration
• Enhanced alert mechanisms
• Documentation and educational resources

📄 License

This project is licensed under the terms specified in the LICENSE file.

🔗 References

• Edge Impulse Documentation
• ESP32 Arduino Core
• MPU-6050 Sensor Library
• PlatformIO Documentation

---

Built for educational and assistive purposes - helping create safer learning environments through technology.