NeuroPulse AI

What I worked on

• ECG signal preprocessing pipeline
• HRV feature extraction and PCA-based feature processing
• Training and comparing SVM, XGBoost, CNN, ResNet, Inception, and Transformer models
• Three-class problem design: normal / pre-seizure / seizure
• Model evaluation methodology
• Alert-workflow concept for early warning

Technologies

• SVM
• XGBoost
• CNN
• ResNet
• Inception
• Transformer
• PCA
• HRV analysis

Purpose

Epileptic seizures strike without warning, which is a large part of what makes them dangerous. Research suggests the autonomic nervous system leaves traces in heart activity before some seizures. The project investigates whether ECG-derived heart-rate-variability (HRV) features can be used to detect a pre-seizure state early enough to raise a warning.

The approach

The system follows a research pipeline: ECG data → signal preprocessing → HRV feature extraction (with PCA and feature processing) → classification → alert workflow. It was designed around a three-class problem — normal, pre-seizure, and seizure — and compared classical and deep approaches side by side: SVM and XGBoost on engineered features, and CNN, ResNet, Inception, and Transformer architectures on richer representations.

Key features

• End-to-end ECG → warning research pipeline
• HRV feature engineering
• Classical ML and deep-learning comparison
• Three-class seizure-state classification design
• Early-warning alert workflow concept