Research 2026: Noninvasive Hypoglycemia Detection and TinyML on the Edge

My current research pulls together the two threads that run through everything above — embedded systems and machine learning — and points them at wearable health. Two papers are in progress for 2026.

Noninvasive hypoglycemia detection from smartwatch signals

A deep-learning framework for real-time hypoglycemia detection using heart-rate dynamics captured by a consumer smartwatch. The goal is to flag dangerous low-glucose events without a finger-prick or invasive sensor, combining classical signal processing with edge-AI inference so detection can run on the device itself.

TinyML and federated learning for IoT health monitoring

A privacy-preserving workflow that trains personalized glucose-risk models directly on low-power IoT devices. Using TinyML and federated learning, each device improves its own model locally and shares only model updates — never raw health data — so personalization doesn't come at the cost of privacy.

Where it comes from — the thesis

The through-line goes back to my undergraduate thesis at BUET (2014): active noise control inside a duct using an analog circuit, using destructive interference to attenuate noise by around 20 dB, supervised by Professor Dr. Mahbubur Razzaque of the Department of Mechanical Engineering. Real-time sensing and control on constrained hardware has been the constant ever since.

---

Both papers are targeted for 2026 — I'll post updates here as they progress.