A number of technological advances in recent years have led to the development of sensors that are tiny, and sometimes even flexible or stretchable. These sensors are ideal for use in wearable devices that continually collect measurements of physiological parameters. Such on-body sensing systems have important applications in the medical field, where this type of data is typically only collected intermittently, and in clinical settings. Always-on wearable sensors have real potential to diagnose health problems long before they might otherwise be detected, leading to an earlier initiation of treatment plans, and better long-term outcomes for patients. But before we get ahead of ourselves, we have to remember that sensing is only one part of the equation.
After all, what good is sensor data without processing units that can acquire and interpret it? These processing units also need to be well suited to wearable platforms, which means they need to be small, flexible, and stretchy. To run cutting edge algorithms, these devices additionally need to be powerful, yet since they will be running on the power of a small, on-body battery, they also need to be highly energy efficient. That is quite a big ask, so it is no wonder that processors meeting this description are exceedingly hard to come by. A collaboration between Argonne National Laboratory and the University of Chicago may one day help to fill this gap, although the work is still in the early stages. They have developed a soft, stretchy, wearable neuromorphic computing chip that excels at running advanced machine learning algorithms, and slowly sips energy.
The chip is constructed on a thin, plastic semiconductor film. Stretchable gold nanowire electrodes are deposited on the film. The chip can be stretched to twice its normal size without causing damage or degrading performance. Using this platform, the team created an array of transistors. The transistors were arranged to perform vector-matrix multiplications, which are a core component of running a neural network. Importantly, this chip provides desirable neuromorphic metrics, including linear symmetric weight updates and good state retention, which are essential for high computing efficiency. The unit also appears to be durable, with the transistors having been proven to stand up to more than 100 million state changes.
To test the real world utility of the prototype chip, the researchers designed and built an implementation of a wearable heart monitor. They paired their neuromorphic processor with a sensor that records electrical signals from the heart. They trained the device to recognize the electrical signals characteristic of a healthy heart, and also those of four pathological conditions. Testing revealed that their chip was 95% accurate in diagnosing the condition of the heart.
Leveraging the Advanced Photon Source at Argonne National Laboratory has helped the team greatly in developing their flexible plastic semiconductor film. The intense X-ray beam emitted by this machine reveals in detail how the molecules that compose the material deform and reorganize when it is stretched far beyond its normal size. A planned upgrade of the Advanced Photon Source will amplify the brightness of the beam it generates by 500 times, which the team believes will help them to further refine their technology in the future.
The principal investigator in the study, Sihong Wang, acknowledged that the technology still needs “further development on several fronts.” However, looking ahead, he believes that “our device could one day be a game changer in which everyone can get their health status in a much more effective and frequent way.”