Artificial intelligence can identify subtle patterns in data, which is particularly useful in medicine. So far, these have been offline processes—doctors perform a medical test, and data from the test is run through a software program after. A real-time process could allow doctors to identify and treat a medical problem much more quickly. One way to detect these patterns in real time would be with an AI system implanted in the body.
In a new study led by researchers from TU Dresden, researchers created a system made from networks of tiny polymer fibers that, when submerged in a solution meant to replicate the inside of the human body, function as organic transistors. These networks can detect and classify abnormal electrical signals in the body. To test their system, the researchers used it to identify patterns in types of irregular heartbeats. Technology like this could be used to detect medical concerns like irregular heartbeats and others, such as high blood sugar.
"What we have demonstrated is a general concept," said Matteo Cucchi, PhD student at TU Dresden and the study's lead author. "It's a general approach that then can be specialized for one particular application."
Neuron-like transistors
To create biocompatible hardware, Cucchi and his colleagues used networks of polymer fibers made out of a carbon-based material called PEDOT. The tiny networks of branching fibers are visible with a microscope. Cucchi and fellow researchers led by Karl Leo, senior author of the study and director of the Dresden Integrated Center for Applied Physics and Photonic Material, where this research took place, were struck by how similar they looked to neurons.
When immersed in an electrolyte (a salt solution) that mimics conditions inside the human body, the networks of fibers become organic electrochemical transistors (OECTs), which, like silicon-based transistors in traditional computers, act as switches for electrical current, though using a different mechanism.
In a traditional silicon transistor, a metal contact controls whether the transistor is on or off. An OECT "works very differently because you contact the channel with the electrolyte, and you change the potential of the electrolyte," said Leo. "In this way, you can control the number of ions which are in the polymer [fibers] or the electrolyte. And that is changing the conductivity." These organic transistors transform electrical inputs into nonlinear signals, like the binary code that computers use, making it usable for computation.
The researchers used an approach to machine learning called reservoir computing for their system. Unlike the highly-structured organization of other machine learning systems, the components are configured randomly to form a reservoir. In the study, the OECTs were random because of the way they were made. The researchers used a method called AC electropolymerization, which involves running alternating current between electrodes across a liquid precursor to PEDOT. Material starts to condense on one electrode and a fiber eventually grows to the other. The process produces fibers with varying resistances and response times, which help transform the electrical inputs into nonlinear outputs.
The researchers input data in the form of electrical signals, replicating the type of ionic information that the system would receive if it were inside the body. The system worked best when the data was encoded into electrical frequencies. The signals are changed and transformed by the "black box" of the reservoir. Then, the researchers could train a smart system to interpret the results as one of several electrical patterns. In this way, the reservoir is trained to recognize and classify patterns of electrochemical information.
Testing the system
One of the datasets the researchers tested their system on was data that represented four types of heartbeats—one normal and three irregular. The AI could correctly distinguish between the four types of heartbeats 88% of the time. Importantly, the heartbeat data was part of an already existing dataset and was not collected from any people as part of the study.
In the future, implantable devices using more specialized versions of this technology might be able to detect unusual electrical signals and medical concerns from within a person's body. The researchers write that this could be particularly useful after surgery. Leo imagines a device with a simple light display that would stay green if a heartbeat were normal and turn red if it became irregular. Cucchi said that such a device could enable doctors to "act immediately on the signal without losing time and money on analysis and invasive procedures."
For now, the researchers said, the technology is nowhere near being used inside a person. The study only examined how a system like this could work. Use of it as a medical implant would require extensive preclinical and clinical testing. The hardware in the study also used outside power and had no internal power source, as an implant would.
The technology also raises questions about the implications of implanting an AI device in a person's body. The authors suggest that this system could be used online, or in real-time, which would raise questions about how that data is presented and collected. Leo says these are important questions to consider alongside future research.
"There is an ongoing discussion about AI and how you apply it, and its potential for misuse," said Leo. "It's definitely an issue here."
Rebecca Sohn is a freelance science journalist. Her work has appeared in Live Science, Slate, and Popular Science, among others. She has been an intern at STAT and at CalMatters, as well as a science fellow at Mashable.
