Seizure Seeker

Seizure Seeker is a browser based EEG event detector; built by multiple students; designed for research and learning. The algorithm runs fully on the client; it reads .edf and .csv EEG files with many channels; computes simple energy features with overlapping RMS windows; finds candidate segments using a robust baseline; and can fuse detections across channels using a K of N vote. Your data stays on your machine.

---

Table of Contents

1. Overview
2. Quick Start
3. Features
4. Usage
5. Repository Structure
6. How the algorithm works
7. Parameter Tuning
8. Performance
9. Customization
10. Roadmap
11. Contributing
12. License
13. Contact

---

Overview

Seizure Seeker performs detection per channel; then it can combine results across channels if you ask it to. Everything happens in the browser; nothing is uploaded. The goal is to help you flag time ranges that deserve a closer look; not to diagnose.

---

Quick Start

1. Serve index.html with a simple static server.
2. Load an EEG file; .edf or .csv both work.
3. For .csv files; enter the sampling rate in Hz.
4. Adjust sensitivity and optional fusion.
5. Run the analysis; inspect per channel and fused segments; export to CSV when ready.

Tip; try examples/sample.edf for a fast check.

---

Features

• EDF and CSV input with many channels
• Per channel RMS energy with overlap
• Optional cross channel fusion with K of N voting
• Controls for Z score; minimum duration; merge gap; and K
• Instant feedback in the page
• Privacy; processing stays local

---

Usage

1. Select a file; choose .edf or .csv.
2. If you chose .csv; enter the sampling rate in Hz.
3. Tune the controls; Z score; Min Duration; Merge Gap; K; and toggle Fusion if needed.
4. Press Analyze.
5. Review the table; you will see per channel rows and fused rows if fusion is on.
6. Click Download as CSV to save the segments.

---

Repository Structure

/
├─ index.html         # Main UI and detection logic
├─ README.md          # This file

---

How the algorithm works

Per channel windowing
The signal is split into 1 s windows with 0.5 s steps; this overlap helps catch edges.

Per channel energy
For each window the algorithm computes RMS; a compact proxy for signal amplitude.

Robust baseline
For each channel the algorithm sorts window RMS values; takes the quietest half; then computes the median and the median absolute deviation MAD. These two numbers resist outliers and artifacts.

Thresholding
For each channel the threshold is median + z * MAD; frames above the threshold are candidates. With Auto Z; the algorithm picks z so that only a small fraction of frames exceed the threshold; you control this with Aggressiveness.

Segmentation and merge
Consecutive candidate frames form runs; runs separated by at most Merge Gap are merged; segments shorter than Min Duration are dropped.

Optional fusion K of N
The algorithm converts per channel runs to a shared time grid; counts votes across channels; and marks time points as positive when at least K channels agree; then it converts those positive points back into segments with the same merge and duration rules.

---

Parameter Tuning

• Z score; higher values mean stricter detection; fewer segments pass
• Min Duration in seconds; removes very short blips after merging
• Merge Gap in seconds; glues nearby runs when the quiet gap is small
• Fusion; enable K of N voting to reduce single channel artifacts
• K; suggested default is about twenty percent of channels with a floor of two
• Sampling rate for CSV; must match the file; EDF rates are read automatically

Practical advice; if you see nothing; lower Z or reduce Min Duration or switch Auto Z to Sensitive. If you see too much; raise Z or increase Min Duration or pick Conservative; and with many channels try enabling fusion or increasing K.

---

Performance

• Complexity; O(N * channels) for RMS and segmentation
• Memory; about O(N/2 * channels) for RMS arrays
• Benchmark; around 0.1 s for a 1 hour; 256 Hz; 16 channel recording on a modern laptop

Numbers will vary with hardware; file size; and browser.

---

Customization

Edit the detect() function in index.html if you want to change any of the following

• window size and hop
• baseline estimation
• merge rules
• fusion logic and voting
• channel selection

---

Roadmap

• Multi channel detection and per lead thresholds
• Frequency band filtering
• Interactive plot overlay
• Command line and API wrapper
• Real time streaming mode

---

Contributing

1. Fork the repo
2. Create a branch
3. Submit a PR
4. Celebrate

---

License

This project uses the CHB MIT Scalp EEG Database from PhysioNet; available under the PhysioNet Credentialed Health Data License. Use is limited to non commercial research purposes.

---

Contact

Questions or feedback; email sarthaktayal2@gmail.com