JOVANA
Explore Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

Why Implants Go Deaf: The Chronic-Stability Problem

An array that works beautifully on day one can lose most of its usable channels within a year — this guide frames the four failure fronts the rest of the track dissects.

From an acute demo to the chronic reality

Volume I sold you the promise: implant a Utah array, sort spikes, decode intent. In an acute experiment — hours to weeks after implantation — that promise largely holds. The hard, unglamorous truth of clinical BCI is that recording once is easy and recording for years is a materials-science and biology problem that nobody has fully solved. A device that must serve a person with paralysis for a decade cannot afford to quietly go deaf after eighteen months.

The phenomenon has a name — chronic signal degradation — and a shape. Track the number of electrodes still returning well-isolated single units over time and you almost always see a decline: fast in the first weeks, then a slower attrition. The engineering goal of chronic recording longevity is to flatten that curve.

The yield-decay curve

The single most useful summary statistic is yield: the fraction of channels still recording usable signal (single units, or often just a signal-to-noise ratio above threshold). A crude but useful phenomenological model treats channel survival like a decay process with a characteristic time constant \tau.

N(t) = N_0\, e^{-t/\tau}, \qquad \mathrm{yield}(t) = \frac{N(t)}{N_0}

A first-order model: the number of viable channels decays from N_0 with time constant τ (empirically months to a few years). Reality is rarely a single exponential — but τ is a handy way to compare devices.

Model the count of still-working channels as decaying exponentially from its starting value, with a time constant \tau (empirically months to a few years) marking how fast. Reality is rarely a single clean exponential, but \tau is a handy one-number way to compare device longevity.

N(t)
The number of viable (working) channels at time t.
N_0
The channel count you started with.
\tau
The decay time constant — how quickly channels are lost.
\mathrm{yield}(t)
The fraction of channels still working, N(t)/N_0.

After one time-constant \tau, about 37% of the original channels remain.

Be careful what you count. Yield defined as 'any threshold crossing' decays far more slowly than yield defined as 'a well-isolated single unit fit for spike sorting'. Many reports of arrays lasting years are counting the former; a decoder that needs stable single units may already be starving. Whenever you read a longevity claim, ask what quality bar was used.

Four failure fronts

Degradation is not one mechanism but four fronts pushing at once. (1) Biological: the body encapsulates and kills neurons near the electrode — the foreign-body response. (2) Mechanical: a stiff probe and soft brain grind against each other with every heartbeat — modulus mismatch and micromotion. (3) Material: coatings and metals corrode, delaminate and dissolve. (4) Encapsulation: water and ions seep into the electronics and shift or short the signal.

The biological front, on a timeline: within hours microglia converge on the insertion wound; over days to weeks astrocytes wall the probe off in a glial sheath, neurons retreat, impedance climbs and spike amplitude falls.

These fronts are coupled. Mechanical micromotion re-injures tissue and keeps the biological response inflamed; a corroded coating raises impedance just as the glial sheath does. That coupling is why single-cause fixes disappoint and why the failure taxonomy in the final guide matters.

Why this is the bottleneck

For a research demo, six good months is plenty. For a product a person lives with, chronic stability is arguably the single biggest gate between today's spectacular lab BCIs and a durable therapy. There are two philosophies for getting through that gate. Fix the interface — softer probes, better coatings, tighter seals — so the neurons and signals last. Or adapt around drift — accept that the signal changes and build decoders that track it. The rest of this track walks the first path through guides 2–4, then hands the second path its due in guide 5.