A skin-like computing patch could give wearable health devices something they have long lacked, instant judgment. By running AI directly on the body in milliseconds, the stretchable system sidesteps server delays and points toward faster responses when every heartbeat matters.

A stretchable computing patch that clings to skin like a bandage may push wearable medicine into much faster territory. Instead of simply collecting data and sending it elsewhere for analysis, the device can process information right where it touches the body, and do it in milliseconds.

That difference matters most when time is thin. In dangerous heart rhythm emergencies, even a short delay can make the gap between useful action and missed opportunity.

The patch was developed by researchers at the University of Chicago Pritzker School of Molecular Engineering, working with scientists at Argonne National Laboratory. In tests, the system used built-in artificial intelligence to analyze several kinds of health data while bent and stretched, without depending on a wireless trip to an outside computer.

“The future that we’re trying to realize is to make wearable and implantable devices smarter,” said Sihong Wang, an associate professor of molecular engineering at UChicago PME and co-senior author of the study, published in Nature Electronics. “It’s helping people have a personal, instantaneous doctor integrated into their devices.”

Most current wearables can sense. They track heart rate, motion, or other body signals, then pass that information along for processing. The analysis usually happens off the body, often on a remote server or conventional computer chip.

That setup works for many consumer uses, but it creates a bottleneck for situations that need real-time decisions. Sending data away and waiting for a result adds delay, complexity, and energy costs.

The new patch tries to solve that by moving computing to the edge, essentially the point where sensing happens. To do that in a soft, body-compatible format, the team built large arrays of organic electrochemical transistors, or OECTs, on a stretchable surface.

These transistors do not behave like the ones in ordinary rigid chips. They process information using both electrical current and ions moving through a gel electrolyte. That gives each device a memory-like quality, allowing it to hold a numerical value over time in a way that resembles how synapses strengthen or weaken in the brain.

That property makes OECTs attractive for neuromorphic computing, a style of computation designed to mimic some features of neural networks. But turning them into dense, stretchable arrays has been a stubborn manufacturing problem.

The soft substrate does not tolerate standard chip-making conditions well, especially harsh solvents or high temperatures. The gel electrolyte brings another headache, because it can spread and merge with nearby parts, causing short circuits.

“What we had to ask was whether we could use or change the properties of these polymers to make them compatible with photolithography, the main patterning method used in the microelectronics industry,” Wang said.

The team’s answer was a new polymer gel that hardens into precise shapes under ultraviolet light. With that approach, they built arrays with up to 10,000 transistors per square centimeter. In a representative 10-by-10 array, they reported a 100% fabrication yield, with device-to-device variation within plus or minus 10%.

“As computer scientists, we’re used to thinking of a neural network weight as just a number,” said Zixuan Zhao, a graduate student at UChicago CS and co-first author of the study. “In hardware, it’s a material, with variability, history, and physical limits. The challenge was to hold those constraints in mind and still compute with enough precision to matter.”

The group tested the arrays on several machine-learning tasks, including one tied to a life-threatening cardiac emergency.

Ventricular fibrillation is a chaotic rhythm in which the heart’s electrical activity breaks down. It is often fatal within minutes. Current implantable cardioverter defibrillators usually respond with a strong shock delivered broadly across the heart, a treatment that can be effective but is also painful and blunt.

Researchers have been exploring a more targeted approach, one that maps the abnormal electrical wavefronts racing across the heart and interrupts them with smaller, more precise pulses. The obstacle has been speed. Those wavefronts move so fast that the analysis has to happen within milliseconds.

“This is a situation where it’s not feasible to have remote computing. It just takes too long,” said Wang. “But if you have a computing device that can do the analysis within the body, it could be possible.”

Using cardiac mapping data from a deidentified donor human heart, the stretchable array identified propagating wavefront locations with a match rate above 99.6% compared with software-based calculations. The system kept that accuracy even when stretched to 60% strain, and related tests showed negligible changes after repeated stretching to 100% strain for 100 cycles.

The researchers also reported that each inference in a simpler classification task took 10 milliseconds.

The same platform was tested on a second medical task, estimating heart attack risk from a set of health measures. In that demonstration, the neural network analyzed nine inputs: age, gender, cholesterol, fasting blood sugar, maximum heart rate, exercise-induced angina, resting ECG, and two ECG ST-segment measures known as old peak and slope.

Across the full dataset of 303 groups, the hardware system reached an average inference accuracy of 83.5%. On the held-out 20% test subset, accuracy was 78.7%. The authors noted that performance could likely improve with a larger training dataset and more neurons in the hidden layer.

The study also showed the array could handle other edge-computing jobs, including feature extraction and data compression through a denoising autoencoder. Beyond medicine, the team explored whether the same stretchable neuromorphic hardware could support offline reinforcement learning for soft robots navigating maze-like environments.

That part of the work points to a broader idea. A soft machine that moves through tight, shifting spaces, such as earthquake debris, may also need onboard computing that does not rely on a constant external signal.

Still, the clearest case here is medical. The patch is not a finished clinical device, and the heart-attack demonstration was a risk-classification task, not a real-time diagnosis in patients. The ventricular fibrillation result relied on donor-heart mapping data rather than a live implanted system. Those limits matter.

Wang’s lab sees the computing array as one piece of a fuller platform, one that could sense, analyze, communicate, and respond while remaining soft enough to match the body.

His group is now working to pair the array with stretchable sensors and wireless communication components. The aim is a device that does not just gather streams of biological data, but interprets them on site and acts fast enough to matter.

“Instead of sending data away to a remote server, we can begin making sense of it right where life is happening,” said Fangfang Xia, a computer scientist at Argonne National Laboratory and co-senior author of the study.

This work suggests a path toward wearable and implantable devices that do more than monitor. By analyzing data directly on the body, a soft computing patch could reduce lag, cut down on data transfer, and support faster responses in cases where timing is critical, especially cardiac emergencies.

The study also shows that high-density, stretchable electronics can handle machine-learning tasks without losing much performance under strain, which could help future devices fit more naturally against skin or tissue.

For now, the results are still at the hardware demonstration stage, but they point toward medical systems that can sense and interpret signals in the same place, at nearly the same moment.