Dramatically reducing the risk of stroke with neuromorphic hardware

Computer architecture inspired by the human brain – to most people, this sounds rather abstract and removed from everyday life. Far from it: a research team from Fraunhofer IIS and Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) has developed neuromorphic hardware capable of detecting early signs of atrial fibrillation and considerably reducing the risk of stroke. In other words, this application has an entirely practical place in everyday life.

A stroke is often preceded by sporadic atrial fibrillation. But this tends to go unnoticed because health insurance companies will cover the costs of a long-term ECG only for a limited period of a few days. Finding an arrythmia in such a short timespan would be due more to luck than anything else. So a better alternative is to perform the ECG with the help of personal wearables. Fraunhofer IIS has already developed demonstrators for this application: FitnessSHIRT can record a single-channel ECG, while CardioTEXTIL can perform a three-channel ECG of medical quality and raise the alarm if it detects atrial fibrillation.


Innovation competition winner


Such wearables must be as energy-efficient as possible in evaluating the ECG data they collect. With this in mind, Germany’s Federal Ministry of Education and Research (BMBF) launched the “energy-efficient AI system” pilot innovation competition. The challenge was to detect atrial fibrillation at an accuracy of at least 90 percent while consuming as little energy as possible. The competition attracted 27 entries, 11 of which were given the chance to develop their solutions. The latter group included the team from Fraunhofer IIS and Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) with its “Lo3-ML” (Low-power low-memory low-cost ECG signal analysis using ML algorithms) project – and triumphed as one of the four winning teams.  


Each of the participating teams was furnished with 16,000 individual, two-minute ECG recordings by the Charité university hospital in Berlin. Fraunhofer IIS handled the first development stages, specifically the team led by Dr. Marco Breiling and Matthias Struck. Essentially, Struck’s group supplied the medical technology know-how, both teams developed the algorithms and tools, and Breiling’s team took care of the final energy efficiency optimization. One of the main approaches was to use ternary weights. “In neural networks, input signals are weighted using special numerical factors and then added up. But multiplication using several bits is very complex and therefore energy-intensive,” Breiling says. So the team based their approach on ternary weighting – in other words, they weighted the input signals using three values: –1, 0 and +1. Should an intermediate signal be unimportant in a specific neuron, for instance, it is multiplied by zero – producing zero. “We also designed the tools we needed to develop such ternary neural networks,” Struck says.


Sleeping saves energy 


Their partners at FAU, in particular the group led by Dr. Marc Reichenbach, then used the resulting algorithms to design a digital circuit using a hardware description language. In turn, this source code was relayed to the Dresden branch of Fraunhofer IIS, specifically the group led by Dr. Jens Döge. They converted the code into a specification that can be used to fabricate a circuit. In the interest of energy efficiency, the researchers had part of the chip most of the time “go to sleep.” This is because the high-speed digital chip needs just 0.02 seconds to evaluate 12.7 seconds of ECG signal data. While one part of the chip is always active, collecting data, the other part is woken up only when it has work to do – which at 0.2 percent of the time represents an energy saving of up to 95 percent. “Another key point is the use of non-volatile storage,” Döge says. This ensures that the chip’s memory isn’t wiped clean whenever it wakes up, but rather that all the information – like the ternary weights – is available even after power-up. This non-volatile storage is referred to as resistive RAM, or RRAM. The required RRAM read and write circuits were developed at FAU by a team led by Prof. Amelie Hagelauer.


Döge believes that his team’s most important contribution to the project was perhaps starting the design process at a very abstract level. “Normally, everyone has their own idea about where we’re going – but simply throwing these ideas together wasn’t going to get us very far. Instead, we had to start right at the beginning, with the architecture, so that we could get a detailed picture of where we could save most energy. Over the course of the project, this was refined in numerous optimization stages. We simulated the power consumption of the overall system, factoring in the influence of the individual components.” Breiling, who coordinated the project, is thrilled about the collaboration: “I’ve been at Fraunhofer IIS for 19 years and this has been the best and most motivated collaboration I’ve ever experienced.”


Use of the new technology is not limited to the evaluation of ECG data, but can also be adapted to assess all kinds of time series – including those relating to hearing aids and for condition monitoring and predictive maintenance of technical systems, and, of course, further bio signals. The prize the team won in the pilot innovation competition is a follow-up project in which it can apply its results. “We want to ensure that Germany is at the forefront of practical AI applications,” Breiling says. “The follow-up project comes at just the right moment: it allows us to take the development of neuromorphic hardware to the next level as part of our next-generation computing initiative.”

 


Article by Dr. Janine van Ackeren

 

Contact

Your questions and suggestions are always welcome!

Please do not hesitate to email us.

Stay connected

The newsletter

Sign up for the the Fraunhofer IIS Magazine newsletter and get the hotlist topics in one email.

Home page

Back to home page of Fraunhofer IIS Magazine