Source-linked AI summary
EITWatch: Smartwatch-Integrated Planar Electrical Impedance Tomography for Hand Gesture Recognition
Xuanyou Liu, Novel Alam, Karan Ahuja
TL;DR
Prior wrist-EIT systems require electrode coverage beyond the smartwatch case-back patch and separate analog front ends, leaving the patch-only feasibility question open. EITWatch tests it with an eight-electrode planar ring and multi-depth scanning, achieving over 90% within-session accuracy while transfer across sessions and users remains substantially lower.
Problem
Prior wrist-EIT systems require electrode coverage beyond the smartwatch case-back contact patch and separate analog front ends, leaving patch-only gesture recognition unresolved.
Method
EITWatch uses eight planar electrodes in a 31 mm ring, acquires 35 measurements at 48 Hz, and varies source-sink distances and current paths through multi-depth scanning.
Results
Within-session leave-one-round-out accuracy exceeded 90% for both gesture sets across 12 participants, while cross-session and leave-one-user-out transfer was substantially lower.
Takeaways & Limitations
The planar watch-back patch carries gesture-discriminative EIT signal, supporting feasibility while leaving re-donning and person-independent transfer as open challenges.
Takeaways & Limitations
EITWatch is a feasibility prototype rather than a deployment-ready recognizer because transfer accuracies are much lower and robustness factors were not systematically swept.
Abstract
from arXiv · showhide
Wrist Electrical Impedance Tomography (EIT) senses hand gestures from muscle- and tendon-driven impedance changes, but prior wrist-EIT systems require electrode coverage beyond the watch-back contact patch and separate analog front ends. We present EITWatch, the first wrist-EIT system built around smartwatch case-back geometry, asking whether this contact patch alone can support gesture recognition: eight planar electrodes in a 31 mm ring acquire 35 impedance measurements at 48 Hz. Because a planar array cannot encircle the wrist, EITWatch uses multi-depth scanning to sample multiple source-sink distances and current paths; it beat matched adjacent injection by 15.1/10.4 percentage points (macro/micro) across all 12 participants. In a prompted study, within-session leave-one-round-out accuracy reached 91.4%/92.5% (window/trial) for six macro-gestures, and 90.1%/91.5% (window/segment) for five micro-gestures plus relax; window-level cross-session and leave-one-user-out transfer reached 73.2%/70.4% and 63.1%/55.3% (macro/micro).
1 Introduction
EITWatch asks whether a smartwatch case-back contact patch alone can support wrist-EIT gesture recognition. It combines planar hardware with multi-depth scanning and evaluates feasibility across participants, sessions, users, and matched acquisition protocols.
- System concept: EITWatch is the first wrist-EIT system centered on smartwatch case-back geometry, isolating sensing to an eight-electrode, 31 mm planar ring.Its discrete analog front end extends beyond the watch case, so the prototype does not fully miniaturize the electronics.
- Acquisition design: Multi-depth scanning sweeps source-sink separation and direction because a dorsal planar array cannot surround the wrist.The protocol samples multiple characteristic depth scales and angular current paths within each frame.
- Evaluation: Across 12 participants, within-session leave-one-round-out evaluation exceeded 90% window-level accuracy for both gesture sets.These results measure within-session repeatability rather than deployable recognition.
- Evaluation: A matched comparison showed multi-depth scanning outperformed adjacent injection on the same hardware and tasks, while cross-session and cross-user transfer dropped substantially.The evaluation covers within-session, cross-session, and cross-user recognition.
- Resources: The work releases EITWatch schematics, PCB layouts, and firmware to support future planar wrist-EIT research.The release is provided through the project repository.
2 Related Work
Prior wrist-EIT gesture systems generally require circumferential or localized wristband hardware and external analog electronics. EITWatch targets the unresolved smartwatch-scale case-back sensing problem with a planar array.
- Prior wrist-EIT: Prior wrist-EIT systems use circumferential forearm or wrist bands, localized wristband arrays, or flexible placements beyond the watch-back patch.These systems include Tomo, HighRes, EIT-kit, EITPose, EI-Lite, and BandEI.
- Smartwatch gesture input: Most non-EIT smartwatch gesture approaches require additional hardware or active signal emission beyond the watch itself.Examples include EMG, bio-acoustics, ultrasonic and sonar sensing, body-coupled RF, and depth sensing.
- Surface planar EIT: Prior planar EIT establishes one-sided sensing and depth-dependent resolution, but does not address smartwatch-scale planar arrays for gesture recognition.EITWatch applies this sensing geometry to the watch-back contact patch.
3 Planar Sensing on the Watch Back
The watch-back array senses tissue beneath a single surface by shaping current penetration through electrode spacing. Larger source-sink separation reaches farther in a heuristic sense, while impedance modulation supplies gesture-related signal.
- Planar geometry: EITWatch places all eight electrodes on the dorsal wrist within a 31 mm ring to study sensing beneath a compact planar array.The design examines what the watch-back geometry can sense as electrode spacing varies.
- Physical model: The tissue potential follows the quasi-static Laplace equation, with conductivity distribution determining the electric field in the tissue domain.Current density is defined by the conductivity-weighted potential gradient and curves into the tissue.
- Current penetration: A larger injection-pair distance increases the heuristic characteristic depth scale L/2, but current density decays continuously rather than stopping at a boundary.Adjacent spacing of approximately 11.9 mm weights shallower regions more strongly than opposite spacing of approximately 31.0 mm.
- Gesture signal: Gesture recognition uses impedance modulation from muscle contraction against a static background shaped by poorly conducting bone and fat.Reported forearm measurements show resistance increases during isometric finger-flexor contraction, with localized impedance changes near 10% at maximal voluntary contraction.
4 EITWatch System Design
EITWatch integrates eight watch-back electrodes, embedded analog and digital control, and a multi-depth measurement schedule. The schedule varies injection geometry to broaden depth and path coverage, with validation in a tank and on-body ablation.
- Hardware: Eight 2 mm gold-plated electrodes form a 31 mm ring that fits within a 40 mm watch case back.The electrodes press flush against the dorsal wrist through the custom PCB.
- Hardware: The 40 × 60 mm six-layer PCB places the electrode ring and switching beneath the watch while housing the discrete analog front end in a 20 mm extension.This partition preserves the watch-back footprint while keeping the prototype debuggable.
- Control and acquisition: An ESP32-S3 sequences eight electrodes through a 35-channel multi-depth scanning schedule at approximately 48 Hz.The controller streams the 35 scalar features over Wi-Fi for logging and offline classification.
- Measurement configuration: The multi-depth protocol fixes one source, steps the sink across the other seven electrodes, and collects five differential measurements per injection pair.This yields 7 × 5 = 35 measurements while varying injection geometry relative to the adjacent-injection baseline.
- Measurement configuration: Multi-depth scanning samples multiple source-sink distances and angular paths, trading a small reduction in per-frame channel count for broader coverage.A pilot across all source electrodes found less than 1.5 percentage points of LORO-accuracy variation on two participants.
- Validation: Water-tank validation tested anomalies at 2, 5, 10, and 20 mm using identical hardware and repeated measurements for both protocols.Multi-depth scanning retained stronger deep-anomaly contrast than adjacent injection at 10–20 mm, without isolating depth as the cause.
5 Gesture Recognition Study
The study evaluates two prompted gesture vocabularies using fixed watch-back EIT recordings and fold-isolated feature extraction and classification. It includes six static macro-gestures and five transient micro-gestures plus relax.
- Gesture vocabularies: Two prompted vocabularies comprised six static macro-gestures and five transient micro-gestures plus relax.The micro task evaluates gesture discrimination and relax-state rejection in a six-class setting.
- Participants and protocol: Twelve right-handed participants wore EITWatch on the distal dorsal wrist without conductive gel or per-user contact-quality calibration.Participants secured the watch with a silicone strap during approximately 30-minute sessions.
- Participants and protocol: Macro trials presented six randomized held poses across 10 rounds, while micro trials presented five transient gestures across 10 rounds with repeated executions.Macro gestures were held for 2 seconds; micro recordings included consecutive repetitions.
- Processing and evaluation: Raw 35-channel frames were filtered before leave-one-round-out splitting, then converted into 175 window features for ExtraTrees classification.Features comprised five statistics per channel; settings were selected within training folds and fixed before scoring held-out rounds.
- Gesture vocabularies: Figure 5 distinguishes the six static macro classes from the six dynamic micro classes, with arrows marking micro-gesture motion direction.The micro set includes relax.
6 Results
EITWatch achieved high within-session recognition, but accuracy declined under cross-session and cross-user transfer. Multi-depth scanning also outperformed matched adjacent injection on the same device and tasks.
- Within-session recognition: 91.4% macro window and 90.1% micro window accuracy were achieved in within-session leave-one-round-out evaluation.Trial-level macro accuracy was 92.5%, and segment-level micro accuracy was 91.5%.
- Within-session recognition: Figure 6 reports pooled leave-one-round-out confusion matrices for macro trials and micro segments across 12 participants.Cells show counts and row-normalized percentages.
- Transfer evaluation: 73.2% macro and 70.4% micro window accuracy were achieved in 48-hour cross-session transfer without retraining.The model was trained on first-session data and evaluated on complete second-session recordings.
- Transfer evaluation: 63.1% macro and 55.3% micro window accuracy were achieved in leave-one-user-out evaluation.These transfer results were substantially below within-session leave-one-round-out performance.
- Protocol comparison: Multi-depth scanning improved matched adjacent injection by 15.1 percentage points for macro gestures and 10.4 points for micro gestures.The comparison used the same device and pipeline; every participant improved despite fewer channels per frame.
- On-device performance: The on-device measure-to-inference update took 3.07 ms for macro gestures on the ESP32-S3 pipeline.Frames arrived every 20.8 ms at 48 Hz; the reported latency was measured on the on-device path.
7 Limitations and Discussion
The results support discriminative signal in the planar watch-back patch, while transfer, robustness, and deployment scope remain limited. The authors identify calibration, hardware integration, and broader evaluation as next steps.
- Scope and transfer: EITWatch is a feasibility prototype rather than a deployment-ready recognizer because transfer accuracies were much lower than within-session results.The central finding is that the watch-back patch carries gesture-discriminative EIT signal.
- Un tested conditions: Robustness was only indirectly probed because strap force, watch rotation, contact impedance, and wrist tissue composition were not swept.These factors remain untested sources of variation.
- Evaluation scope: The prompted micro task used annotated intervals and a small vocabulary, leaving open-world spotting, false-activation rates, and live UI feedback for future study.The study therefore does not evaluate unconstrained online interaction.
- Future work: Next steps include integrating the front end into a watch case, improving transfer with calibration or user-adaptive training, and broadening gestures and populations.These are proposed directions rather than evaluated outcomes.