A woman in the Netherlands with late-stage amyotrophic lateral sclerosis has been using an implanted brain-computer interface to communicate since 2015. During the day, the system works as intended: she attempts hand movements, the implant decodes them, and a cursor clicks. At night, the same decoder went haywire, registering 245 phantom click-commands and 13 false caregiver calls every hour.
That finding, published this month in Scientific Reports by a team at UMC Utrecht Brain Center led by Sacha Leinders, exposes a practical gap that the BCI field has largely talked around. Most implant studies evaluate performance during structured daytime sessions. Almost none address what happens when the patient falls asleep with the device still running.
The patient — quadriplegic, ventilator-dependent, and in her fifties at the time of implantation — had two Medtronic Resume II electrode strips (four contacts each) placed over her left sensorimotor hand knob, connected to a subcutaneous Activa PC+s amplifier beneath her clavicle. The system had been operational for more than five years when the Utrecht group began systematically recording its overnight behaviour in 2020.
What they found was unambiguous. Power and variance in the high-frequency band were elevated during sleep in 88 percent of the 337 nights they logged. The low-frequency band showed similar shifts. These changes are consistent with known electrophysiology of sleep — cortical neurons fire differently during slow-wave and REM stages — but no one had quantified how badly they break a daytime-trained BCI decoder in a real home-use scenario.
The 245 false activations per hour are not a minor nuisance. Each one represents a click the patient did not intend. In a communication interface, that means garbled text. In a system linked to a caregiver alarm, it means the person who comes running at 3 a.m. eventually stops coming at all.
To address this, the Utrecht team designed a nightmode algorithm that exploits a different signal feature: the low-frequency rebound that follows an attempted movement. Instead of decoding continuous motor imagery as the daytime system does, nightmode requires the patient to perform a deliberate sequence of attempted hand movements timed to her ventilator cycle — one movement per breath, roughly every eight seconds. The system counts correctly timed cycles before triggering an activation, making accidental commands far less likely.
Over approximately 18 months of home use, nightmode operated error-free on 79 percent of nights. Unintended activations dropped to roughly once every 12 nights. The trade-off is speed: the deliberate sequencing means it takes about 90 seconds to complete a command, compared to a few seconds during the day. For a caregiver call at 3 a.m., that delay is acceptable. For fluid communication, it is not.
The study has limitations the authors acknowledge. No sleep staging was performed, so the relationship between specific sleep phases and decoder failure is unclear. Passive limb movements during caregiving routines — turning the patient, adjusting equipment — may have contributed to some false activations. And the findings come from a single participant; whether the same signal drift patterns hold across different implant types and patient populations remains an open question.
Still, the practical implications extend well beyond Utrecht. Neuralink, Synchron, and Paradromics are each building toward devices they intend for long-term, continuous use. Neuralink’s N1 chip samples from 1,024 electrodes in motor cortex. Synchron’s Stentrode sits in the superior sagittal sinus. Both will encounter the same neurophysiological reality: sleep changes brain signals, and decoders trained on waking data will not work at night without adaptation.
The Utrecht group’s contribution is not a breakthrough in the conventional sense. It is a careful, 337-night accounting of what goes wrong when a BCI meets the full 24-hour cycle of human neurology. For an industry racing toward commercial devices that patients will live with around the clock, that accounting is overdue.