Technology

What runs under the hood when NeuroShield scores an EEG window.

Input

(Batch, 1, 7500): roughly 30 seconds of single-channel EEG sampled at 256 Hz.

Conv1D feature trunk

Three blocks progressively downsample time while lifting channels:

• Block 1: Conv1d(1 → 32, k=7) + BatchNorm + ReLU, MaxPool(/2), Dropout 0.2
• Block 2: Conv1d(32 → 64, k=5) + BatchNorm + ReLU, MaxPool(/2), Dropout 0.2
• Block 3: Conv1d(64 → 128, k=3) + BatchNorm + ReLU, MaxPool(/2), Dropout 0.2

Temporal stack

A 2-layer bidirectional LSTM (hidden=64, dropout 0.3) models long-range dynamics and frequency transitions, followed by attention-weighted pooling.

Multi-output head

One shared representation, two calibrated risk scores (0–1): Schizophrenia and Alzheimer’s. Trained on PhysioNet EEG MMI, TUH EEG Corpus, DEED, ADNI and COBRA. Validated accuracy: 94.2%.