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

Closing the Loop: Adaptive and Responsive Neuromodulation

With a target and a biomarker in hand, the engineering question becomes a control problem: when should stimulation turn on, how much, and how do you sense and stimulate in the same brain at the same time?

From continuous to contingent stimulation

Conventional DBS stimulates continuously, whether or not the patient needs it at that moment. Adaptive (closed-loop) DBS instead makes stimulation contingent on the biomarker: deliver it when the brain enters a symptomatic state, ease off when it does not. In psychiatry this often borrows the logic of responsive neurostimulation — a technology proven for epilepsy, where a device detects a pathological pattern and responds — repurposed to detect a mood or symptom biomarker instead of a seizure.

The control policy

The policy π is the rule mapping biomarker to stimulation. The simplest is a thresholded rule with hysteresis: turn on above one threshold, off below another, so the device does not chatter around a single boundary. A gentler alternative is proportional control that scales the dose with how far the biomarker sits from its healthy set-point. Both are biomarker-driven stimulation; the choice trades responsiveness against smoothness and against the risk of driving the loop into oscillation.

u_t = \begin{cases} u_{\max}, & b_t > \theta_{\text{on}} \\[2pt] 0, & b_t < \theta_{\text{off}} \end{cases} \qquad\text{vs.}\qquad u_t = \mathrm{clip}\!\big(K\,(b_t - b^\star),\; 0,\; u_{\max}\big)

Two canonical policies. Left: a hysteresis threshold (θ_on > θ_off prevents chattering). Right: proportional control toward a set-point b★, clipped to the safe stimulation range. Real adaptive-DBS controllers are elaborations of these two ideas.

More ambitious controllers optimise an explicit cost that balances symptom control against stimulation dose. Because side effects and battery both scale with charge delivered, the objective is not 'suppress the biomarker at all costs' but 'suppress it cheaply.'

J = \sum_{t}\Big[\, \underbrace{Q\,(b_t - b^\star)^2}_{\text{symptom cost}} \;+\; \underbrace{R\,u_t^{2}}_{\text{dose / battery}} \;+\; \underbrace{\lambda\,\mathrm{Side}(u_t)}_{\text{side-effect penalty}} \,\Big]

An optimal-control view of adaptive stimulation. Q weights symptom deviation, R penalises charge (side effects and battery), and λ penalises stimulation-induced states you want to avoid. The tuning of Q, R, λ is itself a clinical value judgement, not a purely technical one.

Reading and writing in the same brain

A closed loop must sense while it stimulates, and stimulation is orders of magnitude larger than the signal you want to read. The stimulus contaminates the recording as a huge stimulation artefact; naive amplifiers saturate. Practical systems either interleave sensing and stimulation in time (blanking the amplifier during each pulse) or subtract a template of the artefact — the same read–write coexistence problem you saw in bidirectional interfaces, now with a slow biomarker and chronic timescales.

There is also a subtler, biological read–write coupling: stimulation itself evokes a measurable neural response. Rather than a nuisance, this can become a signal — a stimulation-evoked marker such as evoked resonant neural activity gives a direct read-out of target engagement, letting the device confirm it is actually modulating the intended circuit.

Dose, duty cycle and the hardware that runs it

In the adaptive setting the same target map becomes dynamic: the controller modulates a chosen network node only during symptomatic states, so 'the target' is now a node plus a duty cycle rather than an always-on site.

D = \frac{1}{T}\sum_{t=1}^{T} \mathbb{1}\!\left[\,u_t > 0\,\right], \qquad \bar{P}_{\text{stim}} = D \cdot P_{\text{on}}

The duty cycle D is the fraction of time the stimulator is active; average power (and thus side-effect and battery load) scales with it. A closed loop that fires only during symptomatic states can achieve a low D — the quantitative promise of 'the same with less.'

None of this is hypothetical hardware. Clinically-available and research sensing-enabled pulse generators can chronically record a local field potential and stream it out, which is what made human closed-loop psychiatric studies possible at all. The remaining bottleneck is rarely the electronics — it is trusting the biomarker and the policy enough to let the device act autonomously for months, a theme the final guide takes up under evidence and ethics.