Capstone / ML|2025 - 2026

On-Chip ECG Classification System

A senior capstone ECG classification pipeline designed for wearable deployment under strict power, latency, and memory constraints.

PythonPyTorchGPU trainingSignal processingMachine learningSpiking neural networksTime-series classification

Project Overview

Built the full ECG pipeline from raw signal processing through inference, including filtering, segmentation, feature extraction, model training, and evaluation.

Developed and evaluated multiple learning approaches, including conventional neural networks and spiking neural networks, to study which methods were most promising for low-power wearable hardware.

Used PyTorch with GPU acceleration to train and compare models while tracking confusion matrices, classification metrics, and system-level tradeoffs.

Focused the project on what could realistically run on embedded hardware, balancing model quality against power, memory, and latency constraints.

Technical Details

Signal-processing pipeline included filtering, segmentation, and feature extraction.
Model development in PyTorch with GPU acceleration for training and evaluation.
Compared conventional neural networks with spiking neural network approaches.
Evaluated time-series classification performance with confusion matrices and classification metrics.
Framed around on-device inference rather than cloud processing.

Engineering Challenges

Working with noisy biological signals and building a preprocessing pipeline that preserved useful structure.
Balancing model complexity against realistic hardware limits on memory, latency, and power.
Comparing architectures in a way that reflected deployment constraints instead of isolated benchmark accuracy.
Building an evaluation workflow that stayed meaningful across multiple model variants.