The surgical variable
A decoder assumes a fixed relationship between neurons and electrodes, but a surgeon establishes that relationship one patient at a time. The surgical implantation approach is itself a design choice with a signal/risk tradeoff: an open craniotomy places an intracortical array directly in cortex for the richest signal but the highest surgical burden; an endovascular electrode threads in through a vessel with no craniotomy but coarser signal; minimally invasive and epidural routes sit between. Placement varies by millimetres between surgeons and patients, and those millimetres change which neurons you record and how the decoder must be calibrated.
Safety surveillance
Once a device is in a person, safety becomes a continuous obligation. Adverse-event reporting tracks the expected (transient pain, minor infection) and, crucially, the serious and unanticipated events — device-related infection requiring explant, intracranial haemorrhage, hardware failure, unexpected stimulation effects. A Data Safety Monitoring Board reviews these on a schedule and can pause or stop the trial. Rates are reported per exposure, not per participant, because risk accumulates with implant-time.
\hat{\lambda} \;=\; \frac{k \;\text{events}}{\sum_i t_i \;\;(\text{person-years})}Exposure-adjusted incidence rate; with only a few events k, report an exact Poisson confidence interval rather than a normal approximation.
Turn a raw count of bad events into a fair rate by dividing not by the number of patients but by their total time at risk. Someone implanted for five years contributes more exposure than someone implanted for one, so person-years is the honest denominator. With only a handful of events, use an exact Poisson interval, not a normal approximation. This underpins adverse-event reporting.
- \hat{\lambda}
- The estimated event rate, in events per person-year.
- k
- The number of events observed.
- t_i
- The time person i was actually at risk.
- \sum_i t_i
- Total exposure in person-years — the denominator.
3 infections across 200 person-years is a rate of about 0.015 per person-year, or roughly 1.5 per 100 person-years — but with only three events the confidence interval is wide.
The home-use gap
The widest chasm in the whole track sits between a supervised lab session and independent home use. In the lab an engineer recalibrates a drifting decoder, resets a dropped connection, and coaxes performance. At home there is a person with severe disability and perhaps a family member. Setup time, mounting the system, recalibration triggered by day-to-day signal drift, and every failure mode must be handled by non-experts — or the device is abandoned regardless of how well it decodes. Caregiver burden is a first-order design constraint, not an afterthought.
Human factors: designing for the actual user
Regulators treat human factors and usability as a safety discipline in its own right (guided by standards such as IEC 62366). The relevant user is not the graduate student who built the system; it is a person who may be unable to move, speak, or troubleshoot. Use-error is analysed like any other hazard: what happens if a lead is mis-seated, a calibration is skipped, or a caregiver misreads a status light? A usable interface for this population is often the difference between a device that is worn daily and one that stays in a drawer.
Explant and the duty that outlasts the trial
Devices come out. Explantation follows infection, hardware failure, study completion, or a participant's choice to withdraw — and it carries its own surgical risk and, often, real loss for someone who came to depend on the device. This exposes the field's quiet ethical debt: post-trial obligations and continued access. When a study ends, who maintains the implant, funds its upkeep, or lets a participant keep a capability they have integrated into their life? A responsible protocol answers this before the first implant, not after.