Northwestern Smartwatch Tech Reads Hands for Prosthetics

Your smartwatch knows how many steps you took and how fast your heart beat, but it has almost no idea what your hands are doing. That blind spot matters, because hands are how people act on the world, and for amputees and patients in care settings, tracking hand and arm movement can be clinically meaningful. Engineers at Northwestern University’s SPICE Lab say a standard smartwatch can do far more here than it does today.

According to Northwestern’s November 4, 2026 announcement, the lab has built two systems that read hand and arm activity from the wrist alone: one, called EITWatch, senses the muscles and tendons beneath the skin; the other, TransfHAR, learns new activities from a handful of demonstrations. The researchers say the work could one day help control bionic prosthetic hands and support patient monitoring in places like elder care.

Two projects, one idea: the wrist is always there

Both studies were presented at the Association for Computing Machinery’s Symposium on User Interface Software and Technology (UIST 2026). Karan Ahuja, an assistant professor of computer science at Northwestern’s McCormick School of Engineering who directs the Sensing, Perception, Interactive Computing and Experiences (SPICE) Lab, is principal investigator on both.

“Your hands are how you act on the world, but your smartwatch knows almost nothing about what they’re doing,” Ahuja said in the university’s release. “These two projects show how much the wrist alone can reveal – one by sensing the muscles and tendons beneath the skin, the other by learning how the arm moves.”

The motivation is partly about privacy and practicality. Today, the most reliable way to know what hands are doing is a camera. As Ahuja put it, that is “a privacy nightmare,” and it only works when the hand is in view. A watch is worn all day and needs no line of sight.

EITWatch: looking under the skin with tiny electrical signals

EITWatch uses a technique called Electrical Impedance Tomography (EIT). According to the project’s paper abstract, eight planar electrodes arranged in a 31 mm ring on the back of the watch take 35 impedance measurements at 48 times per second. Tiny electrical signals pass through the wrist, and because muscles and tendons shift with each gesture, every movement leaves a distinct pattern the system can learn.

The engineering challenge is that a watch back cannot wrap around the wrist the way earlier wrist-EIT systems did. The team’s answer is “multi-depth scanning,” which samples several source-to-sink distances and current paths. The paper reports that this approach beat a conventional adjacent-injection method by 15.1 and 10.4 percentage points across 12 participants.

The results are encouraging but early. Northwestern reports that the prototype recognized six common hand gestures with over 92% accuracy and five subtler finger movements with over 91%. The paper is candid about the limits: accuracy dropped to roughly 73% and 70% across sessions, and to about 63% and 55% when tested on people the system had not been calibrated for. That matches the university’s note that the watch works best when tuned to a single wearer.

Co-lead author Novel Alam said the next steps are fitting the electronics inside the watch and removing the calibration step entirely. “Any wearer puts the watch on and the gestures work,” Alam said. Alam also highlighted accessibility: “Touchscreens shut out many people. A second input channel built into the device gives them another way in.”

The prosthetics angle follows from the physics. Because the same wrist signals encode what the hand intends to do, the researchers are exploring whether this kind of sensing could help control prosthetic hands.

TransfHAR: teaching a watch new activities in a minute

The second project tackles a different gap. Most activity recognition is limited to a fixed list of common motions. TransfHAR, co-authored with researchers from Carnegie Mellon University and Google Health Research, uses only the watch’s existing accelerometer and gyroscope.

According to Northwestern, the system was pretrained on more than 2,700 hours of publicly available, unlabeled movement data from everyday activities such as walking and exercise. Users can then name a new activity, demonstrate it, and add it to their personal set. The paper’s abstract reports that, in a lab study with 10 participants each performing seven new wrist activities, the system reached 90.4% balanced accuracy after a single one-minute recording per activity, and 86.7% with five examples. Across three offline cross-dataset tests, it matched or beat fully supervised baselines by 6.2 balanced-accuracy points on average.

The healthcare use case is passive monitoring without cameras. “In elder care, a nurse could passively see whether a patient brushed their teeth or shaved, without a camera in the room,” Ahuja said. The release also mentions tracking repetitions of physical therapy exercises. Ahuja added that the team is scaling up its data “with partners at Google” and expects the approach to carry over to rings and wristbands.

What it means

Prosthetics could gain a cheaper, wearable control signal. Sensing tendon and muscle changes from a watch-sized patch is a far lighter setup than many lab-grade alternatives, though this application is still being explored.
Care monitoring without cameras. Personalized activity recognition could help care teams log routines and therapy exercises while preserving residents’ privacy.
Accessibility gets a new input. Gestures such as a thumbs up could become commands for people who struggle with tiny touchscreens.
Personalization is the hard part. Both systems perform best when tuned to one person. Making them work out of the box for anyone is the key hurdle before products appear.

These are research prototypes, not cleared medical devices, and the announcement makes no claims about treatment or diagnosis.

Conclusion

Northwestern’s SPICE Lab is making a compelling case that the hardware already on millions of wrists is underused. One project peers beneath the skin; the other learns from motion. Together they suggest a future in which a watch understands hands well enough to steer a prosthetic, support a care team, or simply let someone answer a message with a thumbs up. The accuracy numbers are strong for prototypes, and the remaining work, miniaturization and calibration-free use, is clearly mapped.