Design and implementation of an energy-efficient edge AI accelerator for real-time IoT healthcare monitoring

Project Abstract

The rapid growth of IoT in healthcare demands intelligent, on-device processing with minimal latency, preserving patient privacy while delivering real-time insights. This work presents the design and implementation of an energy-efficient edge AI accelerator tailored for medical IoT devices that perform real-time health monitoring, anomaly detection, and decision support at the network edge. The proposed accelerator integrates a custom few-shot precision-tlexible neural processing unit with a tightly coupled data path, memory hierarchy, and power-management engine to optimize throughput-per-watt under constrained battery capacity and form-factor requirements. A heterogeneous architecture combines a reconfigurable tensor core, sparse-computation engine, and specialized neural network operators to exploit typical healthcare workloads, such as wearable electrocardiogram (ECG), photoplethysmography (PPG), and vital sign fusion. To address privacy and bandwidth concerns, the accelerator supports end-to-end on-device inference, secure model loading, and robust quantization-aware training to retain diagnostic accuracy across reduced precision, including int8 and mixed-precision formats. The research adopts a co-design methodology, iterating across hardware architecture, compiler toolchain, and machine learning models. A modular hardware-software stack enables automatic mapping of compact neural networks to the accelerator with optimized memory footprints and dataflows that minimize off-chip traffic. The methodology includes (i) profiling real-world healthcare datasets to identify representative network architectures with low parameter counts and low-latency requirements; (ii) designing energy-efficient memory hierarchies with near-threshold SRAM and non-volatile memory integration to sustain long operation in portable devices; (iii) implementing a low-overhead runtime that streamlines graph partitioning, operator fusion, and dynamic voltage/frequency scaling; (iv) developing a hardware-assisted quantization and pruning pipeline to maintain clinical accuracy while reducing model size by up to 20–80%; (v) constructing a secure boot and model integrity mechanism to prevent tampering and protect patient data; (vi) creating a power-management framework that adapts to workload intensity, temperature, and battery state to maximize usable device lifetime; (vii) validating the system with real-time ECG arrhythmia detection, stress monitoring, and dehydration risk assessment under simulated field conditions; (viii) benchmarking against existing edge accelerators and microcontroller units to demonstrate improvements in energy per inference, latency, and diagnostic fidelity. Experimental results show substantial gains in energy efficiency, achieving up to a X× improvement in milli-wepoch per inference and Y% higher sustained accuracy on critical healthcare metrics compared to baseline edge devices. The accelerator demonstrates robust performance under varying patient motion and environmental noise, with secure, real-time inference enabling timely clinical decisions in remote or resource-constrained settings. This work contributes a holistic blueprint for deploying privacy-preserving, energy-aware edge AI in health monitoring, combining architectural innovation, software tooling, and validated medical-grade performance.

Project Overview

What This Project Is About

---

A practical study that explores how to run smart health tasks on small, local devices instead of sending data to a distant server. It focuses on building a compact, energy-saving computer component (an edge AI accelerator) that can process health data from devices like wearable sensors in real time.

---

---

The Problem It Addresses

---

Many health monitoring systems rely on cloud servers, which can cause delays, consume more energy, and raise privacy concerns. This project aims to enable faster, privacy-friendly analysis directly on devices at the patient’s side, reducing energy use while maintaining accurate results.

---

---

Objectives of the Project

---

1. Design a small, energy-efficient processing unit capable of running simple AI tasks.
2. Integrate the accelerator with common health sensors to process data in real time.
3. Evaluate energy use, speed, and accuracy against a baseline system.
4. Explore methods to protect patient data during local processing.
5. Provide a clear implementation plan and possible future improvements.

---

---

What You Will Do Step by Step

---

1. Study existing edge AI and health monitoring systems to identify bottlenecks.
2. Define a small AI model suitable for on-device execution.
3. Design hardware-software integration for the accelerator.
4. Prototype on a development board and connect health sensors.
5. Measure energy use, latency, and accuracy with real data.
6. Tune the design to balance speed and power consumption.
7. Document the build process and create simple demonstration tests.

---

---

Expected Outcome

A working edge AI accelerator prototype that can process health data locally with lower energy consumption and faster responses, along with a short report on its performance, limitations, and potential real-world impact.