Frequency-encoded eye tracking smart contact lens for human-machine interaction

Wearable flexible devices are transforming human-machine interaction by providing seamless, lightweight interfaces that support functions such as haptic sensing, speech and gesture recognition, and motion capture. Vision dominates human perception, and modern eye tracking can infer intention and cognition by detecting gaze. However, dominant camera-based techniques that rely on pupil center–corneal reflection and NIR illumination suffer from environmental light interference, awkward hardware placement, eyelid/eyelash occlusion, and poor universality in daily scenarios. Electrooculography (EOG) is susceptible to muscle signal interference, has lower accuracy, and raises skin-compatibility and social-acceptance concerns due to visible electrodes. There is an urgent need for wearable, imperceptible eye trackers for broad fields including assistive interactions, brain diagnostics, cognitive science, human factors, consumer research, and driver monitoring. This study proposes a miniature, imperceptible, wireless smart contact lens (SCL) eye tracker using frequency encoding to address these issues and enable robust, accurate eye-machine interaction.

Prior work on smart wearable interfaces spans electronic skins and flexible sensors enabling rich HMI modalities. Eye tracking in consumer devices (e.g., spatial computing headsets) typically uses pupil center–corneal reflection with NIR illumination but is limited by ambient light, occlusions, and hardware ergonomics. EOG-based tracking collects ocular dipole potentials via skin electrodes but suffers from EMG interference and limited accuracy, with comfort and social concerns. Smart contact lenses have been developed for AR image projection, physiological monitoring (intraocular pressure, tear glucose), and therapies (drug delivery, color deficiency correction, corneal repair). Scleral coil–based eye tracking offers high angular resolution and fast response and remains a gold standard but requires wired connections, anesthesia, and bulky room-sized generator coils, limiting usability. These gaps motivate a chip-free, battery-free, wireless SCL platform with improved robustness and user acceptance.

Device design and materials: The SCL integrates four chip-less, passive RF tags (coil-shaped RLC resonators) at the lens periphery, each engineered with distinct structural parameters to yield different resonance frequencies (frequency encoding). A portable sweeping-frequency reader (near-field coil plus vector network analyzer, VNA) mounted on eyeglass frames wirelessly measures return loss (S11) and extracts tag-specific signal amplitudes. Eye rotation changes tag orientation and distance relative to the reader, modulating coupling coefficients and thereby the resonance depths. Eye closure produces a common-mode amplitude drop due to eyelid absorption of the near field, distinguishable from differential eye-movement signals.

Lens fabrication: Flexible tags were fabricated on a polyimide (PI) substrate: PI (5 µm) spin-coat on glass; 100 nm Cu seed deposition; photolithography and Cu electroplating to 8 µm for low resistance/high Q; second PI (5 µm) encapsulation; laser cutting to pattern PI; water-assisted lift-off. Tags were molded in a medical-grade silicone elastomer (MED-6015, NuSil) contact-lens mold (base curve 8.6 mm, diameter 13.8 mm) at 150 °C for 15 min. Surface hydrophilization used oxygen plasma (180 s) followed by immersion in 22.2% w/v PVP to form polymer brushes. For in vivo tests, a commercial lens (Clariti, Coopervision) was attached to the silicone’s inner surface for added corneal safety.

Physicochemical and biocompatibility characterization: Central optical zone (4 mm) achieved 89.3 ± 2.3% transmittance (400–800 nm, n=4). MED-6015 surface contact angle decreased from 110° to 7° after hydrophilization and remained highly hydrophilic over one month (n=5). Hydration caused ~2% downshift in tag operating frequencies with reduced Q; Q was used to estimate hydration and calibrate responses. Cytotoxicity was assessed using HCE-T human corneal epithelial cells: lens extracts (prepared per ISO 10993-5) showed >90% cell viability up to 72 h (n=6), and calcein AM fluorescence imaging revealed similar cell density and intensity across groups. Protein accumulation was modeled with BSA-FITC (5 mg/mL, 2 h, room temperature) and removal with Clear Care solution; fluorescence quantification showed lower protein on the SCL than on a bare commercial lens after disinfection (significant p-values reported).

Response modeling and calibration: A 2D eye-movement model (two orthogonal precision rotation stages) rotated a model eye wearing the SCL over ±30° pitch and yaw while a front reading coil recorded S11. Tag amplitudes were computed by summing 20 sampling points (~16 MHz span) around each resonance. A time-sequential eye tracking algorithm was developed, beginning with an implicit swirling calibration: the user (or model) follows a 2.5-turn spiral on a screen (60 cm distance, 27-inch display), logging tag signals at known gaze coordinates. Thin-plate spline interpolation builds a continuous response model over the screen, inherently accounting for individual visual axis vs geometric axis differences. Real-time gaze estimation uses temporal differencing to suppress common-mode drift and map tag amplitudes to screen coordinates.

Repeatability, robustness, and slippage detection: 4000-cycle pitch and yaw tests (12 h) quantified repeatability. Robustness to ambient light and RF interference was assessed under a 24-inch screen and near active smartphone/Wi-Fi router. Slippage detection leveraged tag frequency shifts when a tag approached the corneoscleral junction; one tag’s resonance increased while others remained stable, indicating wear state and slippage direction.

Applications and control logic: Eye-drawing and writing used the calibrated response model to trace NJU letters and a continuous “snake” pattern; errors were computed versus predetermined coordinate sequences. Eye-command interactions (up/down/left/right/closure) were implemented by thresholding differential signals derived from tag amplitudes, with hold durations (e.g., 0.5 s) for command validation. Demonstrations included: (1) Gluttonous Snake game controlled by directional eye commands; (2) web browsing with page switching (Ctrl+Tab / Ctrl+Shift+Tab analogs), page up/down, and screen capture via sustained eye closure (>1 s); (3) PTZ camera pan/tilt controlled by commands with premeasured thresholds. A portable LibreVNA enabled higher sample rates for PTZ control.

In vivo validation: A New Zealand rabbit wore the SCL (with inner commercial lens). The reader coil and portable VNA (7 Hz sampling) captured eye movements; gaze coordinates (downsampled to 1 Hz) drove a robot vehicle via Bluetooth, with a 6-DOF IMU monitoring vehicle motion. Biocompatibility in vivo included 24-h continuous wear (SCL vs bare commercial lens contralaterally) and 8-h/day wear for 1 week (n=3), with slit-lamp biomicroscopy, fluorescein staining, AS-OCT, histology (H&E), and IR thermography.

• The frequency-encoded, chip-free, battery-free SCL accurately detects eye movements and eye closure via four distinct RF tag resonances and a glasses-mounted sweeping-frequency reader.
• High angular accuracy: orientation error <0.5°, below the central fovea’s ~2° high-definition field.
• Eye-drawing precision (swirling-calibration model):
  • NJU letters (n=204 points): mean error X = -0.08 cm (SD 0.32 cm), Y = 0.30 cm (SD 0.28 cm), within the foveal gaze width (2.1 cm band shown).
  • Snake pattern (n=349 points): mean error X = 0.10 cm (SD 0.26 cm), Y = -0.09 cm (SD 0.20 cm).
  • Fingerprint model yielded smaller errors (combined patterns: SD 0.21 cm for both X and Y); swirling calibration slightly increased error due to model approximation but greatly simplified calibration.
• Repeatability: Over 4000-cycle pitch and yaw rotations (12 h), mean SDs of tag amplitudes were 0.061 and 0.043 (max signal strengths 10.226 and 9.165), indicating stable performance.
• Robustness: Tag responses were consistent under a 24-inch display illumination and near an active smartphone and Wi‑Fi router.
• Eye closure detection: Common-mode amplitude drops across all tags distinguished blinks/closure from differential movement signals; additional amplitude reductions at non-tag frequencies also indicated blinking.
• Slippage monitoring: Tag frequency shifts flagged lens slippage and indicated direction, aiding user wear state adjustments.
• Optics and surface properties: Central 4 mm optical zone had 89.3 ± 2.3% transmittance (400–800 nm, n=4). Hydrophilization reduced contact angle from 110° to 7° and maintained hydrophilicity for 1 month (n=5). Hydration shifted tag frequencies by ~2% with reduced Q.
• Cytotoxicity: HCE‑T viability remained >90% up to 72 h with SCL extracts (n=6), and fluorescence imaging showed similar cell density/intensity to controls, indicating low cytotoxicity.
• Protein fouling: After two cycles of BSA-FITC deposition and cleaning, the SCL showed significantly lower residual protein than a commercial lens (multiple comparisons significant: e.g., after second disinfection SCL vs bare lens p = 0.0000159; other p-values reported for before/after and between-lens comparisons; n=3).
• In vivo safety: In rabbits, 24-h SCL wear showed no corneal injury on slit-lamp, fluorescein, or AS‑OCT; histology revealed no erosion, inflammation, or edema. Slight central corneal thinning occurred after 24-h wear for both SCL and bare lens, attributed to compression. One-week wear (8 h/day, n=3) showed no notable abnormalities vs naked eye. IR imaging showed no ocular heating during operation.
• Functionality demonstrations:
  • Game control: Gluttonous Snake controlled by eye commands; 16 turns executed successfully over 400 s.
  • Web interaction: Page switching, scrolling, and screen capture via eye commands and closure (>1 s).
  • PTZ camera: Natural pan/tilt control via up/down/left/right commands; synchronized field-of-view changes demonstrated.
  • In vivo HMI: Rabbit’s eye movements captured and used to drive a robot vehicle in real time via Bluetooth; trajectories matched eye coordinates.

The study addresses limitations of camera-based and EOG-based eye tracking by introducing a minimally obtrusive, chip-free, battery-free smart contact lens that encodes gaze via frequency-resolved passive RF tags. The time-sequential tracking algorithm with implicit swirling calibration achieves sub-foveal angular accuracy and mitigates common-mode drift, enabling both continuous gaze-based drawing and reliable discrete command input. The system demonstrates robustness to ambient light and typical RF sources, and it tolerates practical variations such as lens rotation angle, individual corneal curvature, and small reader distance changes. Safety is supported by in vitro cytotoxicity, reduced protein fouling compared to a commercial lens, absence of in vivo corneal injury over acute (24 h) and subacute (1 week) wear, and no detectable ocular heating. The ability to distinguish common-mode eye closure from differential movement facilitates richer command vocabularies. Demonstrations across software and hardware devices, and an in vivo rabbit-driven robot vehicle, highlight practical HMI potential. Collectively, the findings validate frequency-encoded SCLs as accurate, robust, and biocompatible eye trackers suited for natural human-machine interaction and potential clinical applications (e.g., 3D torsion tracking, REM monitoring).

This work presents a frequency-encoded, wireless, chip-less smart contact lens that tracks eye movement and closure with sub‑0.5° angular error, enabling natural, accurate eye‑machine interaction. A time‑sequential algorithm with implicit swirling calibration delivers high precision while simplifying calibration. The system exhibits strong repeatability and robustness to ambient light and RF interference and accommodates wearing angle, corneal curvature, and reader distance variations. Safety is supported by low cytotoxicity, reduced protein accumulation, and in vivo rabbit tests showing no corneal damage with acute and week‑long wear. Demonstrations include precise eye‑drawing, game and web control, PTZ camera operation, and in vivo robot driving. Future work will target enhanced transparency and flexibility (e.g., using AgNF/AgNW transparent conductors), structural optimization, co‑design of lens/reader/algorithms for simplified or calibration‑free operation, and integration of additional modules (cameras, sensors) for advanced applications in consumer behavior, virtual social interaction, and medical diagnostics and assessment.

• Calibration is currently required: while the implicit swirling method simplifies setup, it introduces slightly higher errors than a full fingerprint model and shows larger errors near screen borders and between calibration lines.
• The reading hardware employs a near-field coil and VNA (including portable versions); although feasible for eyewear integration, this external reader is required for operation.
• Reported operation ranges were characterized within ±30° pitch/yaw in the model; performance outside this range was not detailed.
• Hydration affects resonant frequency and Q factor (~2% downshift), necessitating calibration/compensation.
• In vivo validation was performed in rabbits; human trials were not reported, which may limit immediate generalizability to human use scenarios.
• Slight corneal thinning after 24-h continuous wear (also seen with bare lenses) suggests attention to wearing regimens and fit is needed, though no injury or inflammation was observed.