NeuroLocate

Predict the 3D origin of epileptic seizures using a Physics-Informed Neural Network with Hodgkin-Huxley constraints. All inference runs locally in your browser.

Sample Patients 10 cases with different seizure origins
Upload EEG CSV with 18 channels at 256 Hz
3D Brain Use your hands and palms to interactively move the 3D brain. Bring your palms apart to zoom in, together to zoom out
AI Model Uses brain thermodynamics to rule out impossible places where the seizure originated, increasing accuracy
Click anywhere outside to begin

NeuroLocate PINN Seizure Detection

Ravjoth Brar
Initializing...
Initializing...

No analysis yet

Hover over "Patients" to choose from 10 example cases,
or upload your own EEG data, then click "Run Analysis".

The PINN outputs 3D MNI coordinates of the
epileptogenic zone, constrained by Hodgkin-Huxley physics.

Raw EEG Channels

18-channel scalp EEG (10-20 bipolar montage). Red line = seizure onset.

Upload Format Guide

Your CSV should have:

  • 18 columns — one per EEG channel (10-20 bipolar montage)
  • 2560 rows — 10 seconds at 256 Hz (shorter files are zero-padded)
  • No header row — just numeric values
  • Channel order: FP1-F7, F7-T7, T7-P7, P7-O1, FP1-F3, F3-C3, C3-P3, P3-O1, FP2-F4, F4-C4, C4-P4, P4-O2, FP2-F8, F8-T8, T8-P8, P8-O2, FZ-CZ, CZ-PZ

Overview

This system predicts the epileptogenic zone (EZ) from scalp EEG data. It outputs exact 3D coordinates in MNI space and a confidence spread, visualised as a glowing hotspot on the interactive brain model.

Physics-Informed Neural Network

Unlike standard neural networks, a PINN is constrained by the laws of physics. The model learns to reject predictions that violate the Hodgkin-Huxley equations governing how neurons fire.

EEG Input
18ch × 2560t
→
Encoder
3×128 Tanh
→
EZ Head
(x,y,z,σ)
Physics Head
V,m,h,n,I

The 4 Hodgkin-Huxley Equations

These equations describe electrical signal propagation through neurons via ion channels:

1. Membrane Voltage
2. Sodium Activation (m)
3. Sodium Inactivation (h)
4. Potassium Activation (n)

Inference Pipeline

When you click "Run Analysis", the following happens in your browser:

1
Normalisation
18-channel EEG is z-score normalised per channel to match training data distribution.
2
Forward Pass
Pre-trained PINN outputs initial 3D coordinates and HH state variables (V, m, h, n) per timestep.
3
Physics Optimization (40 steps)
HH variables are refined via gradient descent, minimising the residual of all 4 Hodgkin-Huxley equations to improve physical plausibility.

Training Data

Trained on the CHB-MIT Scalp EEG Database (PhysioNet) — 5 paediatric patients with annotated seizure onsets. 8x data augmentation (noise, scaling, channel dropout, time shift). EZ ground truth derived from energy-weighted centroid of electrodes.

Limitations

Scalp EEG has ~1-2 cm spatial resolution. The predicted EZ is a probabilistic hotspot, not surgical-precision. This tool is for research and educational purposes only — not a medical device.

Research only — not a medical device. Built with TF.js, Three.js & Hodgkin-Huxley PDE constraints.