Overview
As part of a seven-person group at Imperial College London, I contributed to the design and full EU MDR Technical Documentation of a gaze-controlled powered wheelchair — a Class IIa active medical device targeting individuals with severe upper-limb motor impairment such as high-level spinal cord injury, ALS, or quadriplegia. These users retain voluntary eye movement but are unable to operate conventional joystick or sip-and-puff controls.
The device allows a user to navigate a powered wheelchair hands-free by simply looking at where they want to go. An infrared eye-tracking sensor captures gaze direction at 60 Hz, a co-mounted RGB-D depth camera maps that gaze into real-world floor coordinates, and a control algorithm translates the result into proportional motor commands — all within a 200 ms end-to-end latency budget.
The 100+ page technical file produced covers device description and specification, Instructions for Use (IFU), benefit–risk analysis, clinical evaluation, and a Post-Market Clinical Follow-up (PMCF) plan — all structured to EU MDR 2017/745 requirements.
How It Works
The system relies on three hardware layers working in tight coordination:
-
Eye-tracking glasses — Infrared-based, worn by the user, sampling gaze direction at 60 Hz with ±1° angular accuracy. Calibration maps the 2D gaze vector into a coordinate frame shared with the depth sensor.
-
RGB-D depth camera — Mounted on the wheelchair, it captures real-time depth data over a 0.5–4.0 m operating range. An affine transformation fuses the gaze vector with depth to compute a 3D gaze point in world coordinates. A floor-detection algorithm filters gaze landing on non-navigable surfaces (walls, objects) from gaze on the floor — preventing the classic “Midas Touch” problem of unintended activation.
-
Control algorithm — Two modes were designed and validated: a Continuous Control Field interface mapping gaze position to proportional velocity commands, and a Natural Decoder interface using a probabilistic model of natural gaze behaviour to output discrete motion states (forward, reverse, left, right). The Natural Decoder achieves 15.3 bits/second of information throughput — far exceeding conventional switch-based interfaces.
Design & Engineering
The mechanical design integrates the sensor suite into a standard powered wheelchair platform without compromising usability. Sensor mounting was engineered to a <0.1° angular alignment tolerance between the eye-tracker and depth camera optical axes — critical for calibration stability. The enclosures are rated to 1.0 g vibration (5–500 Hz) to handle the dynamic loads typical of motorized wheelchair operation.
The modular architecture keeps the eye-tracker, depth sensor, and controller independently serviceable — an important consideration for a regulated medical device requiring documented maintenance procedures.
Performance
- 15.3 bits/second information throughput via the Natural Decoder interface — significantly exceeding switch-based and discrete command systems
- ≤200 ms end-to-end latency from gaze acquisition to motor command execution
- ±1° angular gaze estimation accuracy (approximately ±5 cm at 2 m operating distance)
- Floor detection range of 0.5–4.0 m, validated on surfaces inclined up to ±55° from the camera axis
- Maximum indoor speed of ≤3 km/h with fail-safe automatic stop on signal loss or calibration drift
- Dwell-time activation threshold of 200 ms (configurable 100–500 ms) to suppress accidental command triggering
Regulatory Pathway
The device was classified as EU MDR Class IIa under EU 2017/745: an active device delivering mechanical energy to the user, with potential for harm if energy is incorrectly directed (unintended movement, excessive speed, delayed halting). The technical documentation addressed:
- ISO 7176-15 — wheelchair information disclosure and documentation requirements
- ISO 20417 — medical device labelling and legibility
- ISO 15223-1 — standardised medical device symbols
- Regulation (EU) 2021/2226 — electronic format requirements for IFU
Risk management identified key hazards (unintended gaze-driven movement, calibration drift, visual fatigue) and documented control measures including gaze confidence filtering, floor-contingent logic, calibration integrity monitoring, and mandatory healthcare professional assessment before independent use. A structured PMCF plan was produced to support ongoing post-market surveillance.