FLEDGED – Feasibility of Low-energy Embedded Deep-Learning-Models Geared for Edge Devices
Idea and relation to VEDLIoT
iBreve develops self-care solutions that focus on prevention & promote healthy living. It was founded by two former Google employees aiming to create a world free of chronic diseases. iBreve developed a patent pending wearable technology to continuously analyze small variations in respiratory rate, breathing patterns and activity to provide personalized interventions.
The goal of our FLEDGED project is to assess VEDLIoT frameworks and technologies to advance the next-generation wearables for a deployment in healthcare for continuous monitoring. We chose our project name as Fledged itself is defined as raising a young bird until it is ready to leave the nest and become able to fly. Thus very fitting for the challenge of making the next generation healthcare wearables ready to take off.
During the project, we assess several VEDLIoT technology areas, which can tremendously improve wearable systems spanning from infrastructure & tools in the center, to the far edge of the system and beyond.
FLEDGED will contribute to the VEDLIoT ecosystem extension with mainly 2 aspects:
- Our use case extends the VEDLIoT Platform by integrating wearable devices that operate on the ‘edge of the edge’ and validating important VEDLIoT concepts, design processes & architecture.
- Our project combines several important aspects within AIoT contributing to the overall robustness of the VEDLIoT concepts & frameworks.
Objectives
The overall objective of FLEDGED is to assess and advance the next-generation wearables architecture for deep learning deployments in remote health monitoring by using VEDLIoT frameworks and to optimize performance, energy consumption, security & privacy for the monitoring use case.
Approach
Currently, most alternatives use AI computations that happen in the cloud instead of locally on the device. However, when it comes to alert functions like for an epileptic seizure it is crucial to move computation ‘over the edge’ to enable instant alarms. A hurdle for deep learning algorithms on-device is that traditional CPU-based systems are limited when it comes to very high bandwidth & low latency, and the need for very large training data sets. The deployment of FPGAs into wearable systems would enable a very high bandwidth and a better ability to scale for high throughput applications. Further FPGAs can be optimized to handle CPU-intensive networking and security functions efficiently.
For the management of chronic diseases like epilepsy, remote monitoring, and fast accurate alerts, for example predicting the onset or severity of a seizure, are of high importance. Thus, one major technological challenge is to keep low false negatives and low false positives rates. Here, deep learning & AI applications can help to increase accuracy and predictability. However, deep learning requires lots of computing power and energy. Both are often not available or very limited locally on site, even when the patient is in a controlled setting like a hospital, rehabilitation clinic or at home. As soon as mobility is important, e.g. when the patient goes for a walk, it is even harder and the available infrastructure is often not sufficient.
In this context wearable innovations with on-device computation capabilities can be extremely valuable, especially in remote monitoring applications like epileptic seizure tracking. Connecting novel data sources like respiratory patterns through AIoT technologies has the potential to create insights into seizure patterns and even forecast the likelihood for a seizure. Thus providing the potential for people with epilepsy to take fast-acting medications or modify activities in anticipation of an impending seizure.
Expected Impact
The project has direct short-term and long-term impact on iBreve’s technology, competitiveness & the overall organization itself. Foremost, the VEDLIoT project allows us a more effective & efficient development of AI & deep learning components for wearable systems. We expect that the project shortens our time to market by 6-12 months. Secondly, the project gives us a first mover advantage by providing us access to crucial resources like novel VEDLIoT technology, in-depth expertise & mentoring and financial contributions. Further, we expect to liaise with important new partners, stakeholders and value chain integrations within the VEDLIoT consortium and beyond.