Xuanyou Liu
Home
UIST 2026 · Short Paper & Demo

EITWatch

Smartwatch-integrated planar electrical impedance tomography for hand gesture recognition

Xuanyou Liu*, Novel Alam*, Karan Ahuja (*equal contribution)

PDF Video DOI BibTeX

UIST 2026, November 2 to 5, 2026, in Detroit, Michigan, USA

Three-panel overview of EITWatch. Top left: the custom board with eight gold electrodes around a 31 mm circle, annotated with its 40 by 60 mm dimensions. Top right: the board beside a standard smartwatch, its microcontroller module visible. Bottom: a person wearing the prototype performs a pinch gesture above the watch.

A smartwatch presses its case back against your wrist all day. EITWatch turns that patch into a gesture sensor with electrical impedance tomography (EIT): tiny currents pass between electrodes while voltages are measured, and moving muscles and tendons give each gesture its own electrical signature.

Prior wrist-EIT wrapped a band around the whole wrist plus external electronics. EITWatch is the first system built around the watch’s own case back: eight electrodes in a 31 mm ring, 35 measurements at 48 Hz. Is that small patch alone enough to recognise gestures?

92.5% six-gesture recognition within a session
40 mm all sensing fits a standard watch-back footprint
8.6 h continuous sensing on a small 300 mAh battery

Sensing through the watch back

Electrical impedance tomography reads muscle motion from the skin; EITWatch asks whether a flat watch-back patch alone can carry that signal.

What EIT is. Electrical impedance tomography senses what is happening inside the body from its surface: tiny alternating currents are passed between electrodes on the skin, and the resulting voltages are measured. The current does not travel in a straight line; it bows down into the tissue. Muscle conducts electricity noticeably better than fat or bone, and it deforms when the hand moves, so each gesture changes the measured pattern.

VVVVVVVVVV12345678+−+−tissueinject 1→2V 3–4V 4–5V 5–6V 6–7V 7–8inject 2→3V 4–5V 5–6V 6–7V 7–8V 8–1
One pair injects current; the others measure voltage, one pair at a time.
A circumferential EIT band wrapped around a cross-section of tissue with bone and fat labelled, wired to stacked circuit boards; beside it, two circular impedance maps reconstructed from the band's measurements.
Electrodes around the cross-section, and two reconstructed images.
EIT overview diagramDiagram of electrical impedance tomography: an AC signal excites one electrode pair around a body cross-section, boundary voltages are measured, and reconstruction turns them into an image of the interior.

From wristband to watch back. Prior wrist-EIT systems wrap an electrode band around the whole wrist and connect it to external electronics, so the sensing hardware is always larger than the watch itself. We explore whether a flat, planar patch on the watch back alone can support hand gesture recognition.

Original wristband A circumferential electrode band wired around the wrist and connected to a laptop

Turn the electrode ring into a wristband

Extra form factor, a separate sensing board, and a wearing position that is not fixed.

Our watch A person wearing the EITWatch prototype and performing a subtle hand gesture

Must EIT encircle the whole wrist?

A smartwatch contacts only a small patch of skin on one side of the wrist.

Implementation

Eight electrodes under a standard watch case, measuring and classifying on the watch itself.

Eight gold-plated electrodes sit in a 31 mm ring beneath a standard 40 mm watch case. A dual-core microcontroller takes 35 measurements per frame at 48 Hz and runs the gesture model onboard.

Our multi-depth injection protocol

Planar EIT cannot surround the wrist, so it sees less of how muscle changes across hand poses. To compensate, one electrode acts as a fixed source while the return steps through the other seven, varying separation and direction every frame. Neighbouring electrodes keep current in the shallow tissue; wider separations push it deeper. A frame therefore samples roughly 6 to 15 mm of depth, where adjacent injection only reads the shallow band.

Adjacent injectionsource and return are neighboursskinfatmuscleboneabout 6 mmpeak current depthMulti-depth injectionone fixed source, the return steps outwardabout 6 mmdeeperabout 15.5 mmpeak current depth
Schematic layers, not measured thicknesses. In an ablation, multi-depth injection beat adjacent-only injection by 15.1 points (macro) and 10.4 points (micro).

User study

Twelve participants performed prompted gestures; VR hand tracking verified the macro poses. They returned two days later to re-don the watch. Recognition is deliberately lightweight: low-pass filtered signals, sliding windows, 175 simple features, and an ExtraTrees classifier, a randomised variant of the random forest.

The gesture set. Top rows: photographs of six static macro-gestures with the differential impedance map each one produces. Bottom rows: five dynamic micro-gestures plus relax, with blue arrows tracing each motion and the corresponding impedance maps below.

Participants performed two gesture vocabularies: six held poses and five quick, subtle micro-gestures plus a relax class. Within a session, recognition reached 92.5% for the held poses and 91.5% for the micro-gestures.

Recognition across sessions and across people remains difficult: re-donning the watch two days later dropped accuracy to 73.2% and 70.4%, and a brand-new user with no training data managed 63.1% and 55.3%. Closing that gap is the open challenge.

0%25%50%75%100%92.5%91.5%Within a session73.2%70.4%Across sessions63.1%55.3%Across peopleno re-donningre-donned two days laternew user, no training dataSix held posesFive micro-gestures + relax
Where it works, and where it does not. Accuracy is high inside one wearing session and falls once the watch is re-donned or worn by someone new.

Next steps

Cross-session and cross-user recognition remain the bottleneck. Next, we will use representation learning and contrastive learning to learn gesture embeddings that stay stable across donning, days, and people, improving generalizability without a full retrain for every wearer.

92.5% within a session (held poses)
73.2% across sessions, two days later
63.1% across people, no user training

Team

This is research from the SPICE Lab at Northwestern University by:

Publication

EITWatch is published in the proceedings of the ACM Symposium on User Interface Software and Technology (UIST ’26), November 2–5, 2026, in Detroit, Michigan. The paper is available on the ACM Digital Library, and the authors’ print is available here (PDF). Open hardware and firmware (schematics, Gerbers, BOM, and ESP-IDF firmware) are available on GitHub. The release was also covered by Hackster.io, CNX Software, and Borncity.

Xuanyou Liu*, Novel Alam*, and Karan Ahuja. 2026. EITWatch: Smartwatch-Integrated Planar Electrical Impedance Tomography for Hand Gesture Recognition. In Proceedings of the 39th Annual ACM Symposium on User Interface Software and Technology (UIST ’26). ACM, New York, NY, USA, 7 pages. https://dl.acm.org/doi/10.1145/3830398.3830673 (*equal contribution) 
@inproceedings{liu2026eitwatch,
  title={EITWatch: Smartwatch-Integrated Planar Electrical Impedance Tomography for Hand Gesture Recognition},
  author={Liu, Xuanyou and Alam, Novel and Ahuja, Karan},
  booktitle={Proceedings of the 39th Annual ACM Symposium on User Interface Software and Technology},
  year={2026},
  pages={1--7},
  publisher={ACM},
  doi={10.1145/3830398.3830673},
  url={https://dl.acm.org/doi/10.1145/3830398.3830673}
}