ECG Adaptive Offloading

CSE398 Final Project — Network-aware adaptive task offloading for ECG anomaly detection in wearable healthcare systems.

Team: Amanda Fogel, Rafa Mantoan Borges, Netaji Meka

Overview

We compare three architectures for ECG anomaly detection on a unified AWS platform:

1. Cloud-only — IoT Core → Lambda → DynamoDB (high accuracy, network-dependent)
2. Edge-only — Greengrass → local TFLite inference (low latency, offline-capable)
3. Hybrid adaptive — network-aware routing between edge and cloud

All three are evaluated on MIT-BIH under four simulated network conditions (stable / degraded / high-contention / offline).

Ownership

Area	Owner
Cloud pipeline (IoT Core, Lambda, DynamoDB)	Rafa
Edge + hybrid pipelines (Greengrass, decision engine)	Netaji
ML model (Hannun CNN, TFLite export, training)	Amanda

Repo structure

• docs/ — architecture, setup, integration specs
• simulator/ — shared ECG + network simulator
• cloud-lambda/ — cloud inference Lambda
• edge-component/ — Greengrass edge component
• hybrid/ — hybrid decision engine
• model/ — model training + export
• analysis/ — notebooks for results + plots
• scripts/ — helpers (test publishing, cert provisioning)

Getting started

See docs/setup.md for AWS resources and environment setup. See docs/integration.md for integration specs (topics, payload format, endpoint).