Putting Their Skin in the Healthcare Game

When it comes to machine learning, it has often been said that data is king. This is especially true when it comes to wearable medical devices that are designed to capture physiological signals and use them to diagnose health problems. Unfortunately, designing a wearable device that can be comfortably worn around the clock so that it can continually capture data is challenging. And while data may be king, that is not the only important factor in designing such a device — the data also needs to be processed on device so that large volumes of data do not need to be transferred over a wireless network. As the number of biological signals being measured increases, remote processing becomes less practical, and it also comes with a whole host of privacy concerns.

Oftentimes a new device will check one of these two boxes — it is either comfortable for long term wear, or it is capable of processing data on device, but not both. A team of researchers at the University of Chicago has recently published the results of their work on a device that can both be worn comfortably and process biological signals on the device. They used polymers to design a soft, stretchy smart skin. Those polymers were also used to create electrochemical transistors that served as the building blocks for a neuromorphic computing chip.

Neuromorphic chips are designed to function more like a human brain than a traditional computing architecture, which makes them ideal for running machine learning algorithms. In particular, this chip was shown to be proficient at vector-matrix multiplication — a critical component of many modern machine learning algorithms. Analysis and storage of data happen in an integrated way that speeds up processing times, and the chip can be optimized to perform well on certain types of problems. Leveraging these characteristics of the device, the researchers designed a wearable smart skin that can monitor heart health.

The monitor was trained to analyze and interpret electrocardiogram (ECG) data. By examining the electrical activity of the human heart, it has learned to classify ECGs as either healthy, or as one of four abnormal states. After training the device, it was tested on new data, and it was found to be able to accurately perform classifications even if it was stretched or bent in any number of ways.

This project is still confined to the laboratory at this time — additional validation will be needed to understand just how effective the device is at recognizing disease signatures. The team does believe that in the future it will be used to send either patients or physicians alerts about potential medical problems. They also see an opportunity for the automatic adjustment of many types of medication doses, similar to the way that insulin can be automatically adjusted today via implanted pumps.

New iterations of the device are already being planned, with adjustments to both the device and the machine learning algorithm being tested. There is much work yet to be done, but this research could lay the foundation for significant advances in wearable medical devices and beyond.