Twenty neurosurgeons tested an ear-worn controller that operates three microscope functions through deliberate tooth clicks while both hands remain on surgical instruments. In a simulated operating-room task, they reported more direct control than when asking an assistant, although the assistant completed the task faster. The German study was posted as an unreviewed preprint on September 11.
Microscope surgery leaves few convenient ways to issue a secondary command. The surgeon's hands are occupied, external controls must remain compatible with sterile work and the eyes need to stay on the operative field. Verbal requests solve the hand problem, but each action then depends on another person hearing, interpreting and carrying out the command.
NeuroClick detects vibrations from an earpiece in the surgeon's right ear. A small neural network distinguishes a tooth click from speech, head movement and other noise. A phone groups the clicks into three patterns for fluorescence, autofocus and image capture, then small actuators press the matching microscope buttons. A tone confirms that the command was received.
The click detector was trained with data from 12 people under quiet and simulated operating-room conditions. The user study then asked 20 neurosurgeons to remove fluorescent targets from a gelatin model using both NeuroClick and an experienced assistant. The order was varied so that practice with the task would not consistently favour one method.
The detector correctly separated the click patterns and background activity in 98.6% of the balanced test score used by the researchers. Surgeons gave NeuroClick a median autonomy rating of four out of five, compared with two for delegation. Median completion time was 146.5 seconds with the device and 136.5 seconds with the assistant, while both methods produced a median of zero task errors.
The user study remains limited to a simulated task at one hospital, and the prototype did not preserve a complete event-by-event record of every command. Mechanical button pressers added delay, and users had little time to practise an unfamiliar gesture. Patient-care trials would require native microscope integration and longer measurements of missed commands, accidental activation, fatigue, comfort and response time.
Related coverage
- Explainable AI Raises Doctors' Lung-Cancer Forecast Sensitivity from 0.72 to 0.87. A clinical decision-support study measures how an explanation interface changes doctors' forecasts.
Sources
- Jonas Hummel and colleagues, arXiv, September 11, 2026. Abstract, authorship, submission date, principal results and licence.
- Full paper. Recognition pipeline, experimental conditions, statistical comparisons and limitations.
- Lead image: Figure 1 from Hummel and colleagues, reproduced without alteration from the paper under CC BY 4.0. It shows the simulated resection experiment, not patient care.