Durable Settings, Not Peak Settings, for Personalized Neurostimulation

Brain stimulation therapies are tuned the way engineers tune network protocols. Someone searches for the knob setting (current, pulse count, timing) that gives the best response today, then reuses it. Networking has shown repeatedly that this breaks. A setting tuned at its peak often sits right next to a cliff, and when conditions drift it falls off. Neurostimulation has the same structure. Responses follow an inverted-U in current intensity, 97.4% of recorded cortical channels respond differently depending on ongoing brain state, and the brain itself drifts from session to session. We propose that one under-examined reason stimulation benefits wash out is simple: the personalized "best" setting, chosen from acute sessions, stops being the best setting within days. Each later session then delivers a weaker effective dose, and the gains never build up enough to consolidate.

The core idea fits in two sentences. When tuning stimulation, choose the setting that is predicted to remain good at future sessions under realistic drift, not the one that scores highest today. This favors flat, robust regions of the response surface over sharp peaks. We then test whether this "durable-setting" rule matches peak tuning in the short term while producing more benefit that persists after treatment stops.

Core Claim

Personalized Bayesian optimization is Gaussian-process search that fits a smooth surrogate to observed responses and picks the next setting to try. Personalized home based neurostimulation via AI optimization augments sustained attention used it to tune transcranial random noise stimulation (weak noisy current through the scalp, 0.1–1.6 mA) from baseline performance and head size. The tuned dose beat both sham and a fixed 1.5 mA dose for low-baseline performers (p = 0.003). That study measured only acute effects and treated the optimum as a fixed point estimate. It also excluded 27% of participants for volatile baselines, which is itself a sign that the response surface moves.

Our claim is that the sharpness of the response surface at the chosen setting predicts how fast that setting's benefit decays across sessions. Sharpness here means how quickly the response falls as the knob moves away. If this holds, an objective that trades a little peak response for robustness to drift should yield larger cumulative effective dose. By the dose-accumulation account of durability, it should also yield more retention after cessation. That account, discussed by Quentin Moreau et al., holds that benefits need sustained, cumulative stimulation to consolidate rather than washing out like an elastic band.

Concretely, let $f_t(\theta)$ be a patient's response to setting $\theta$ at session $t$. We model $f$ with a Gaussian process whose kernel spans both setting and session time, so it learns how quickly the surface drifts. At session $t$ we deliver

$$\theta_t^\ast = \arg\max_\theta \; Q_\alpha\!\left[\, \tfrac{1}{H}\textstyle\sum_{h=1}^{H} f_{t+h}(\theta) \,\middle|\, \text{data up to } t \right],$$

where $Q_\alpha$ is the lower $\alpha$-quantile of the model's predictive distribution and $H$ is the planning horizon in sessions. Peak tuning is the special case $H=0$, $\alpha=0.5$ with no time kernel. The quantile over future sessions is what penalizes sharp peaks. A narrow optimum becomes very uncertain once the surface is allowed to shift, so its lower quantile drops.

How the prior work shapes this

Three networking and dynamical-systems papers supply the intuition that peaks sit near cliffs. Hybrid Control of Bifurcation and Chaos in Dynamic Model of Compound TCP under RED shows a router queue that tips into oscillation and chaos past a critical averaging weight. There, widening the stable margin (from $w=0.269$ to $w=0.336$) mattered more than finding the best-performing point. TCP congestion control for wireless datacenters and BBRv2+ document constants that are optimal in one environment and brittle in others. We borrow the lesson that operating margin should be an explicit design target. We do not borrow the controllers themselves, which assume a known model we lack for the brain.

Parameter Inference with Bifurcation Diagrams tunes model parameters directly against where qualitative transitions occur. Its Jacobian computations scale quadratically with state dimension, and it handles only static bifurcations. Both limits rule it out for high-dimensional, drifting neural recordings, which is why we work with a black-box response surface instead.

On the neural side, Temporal Basis Function Models for Closed-Loop Neural Stimulation establishes state-dependence and provides a fast predictive model of stimulation response. Its evaluation is within-session, so it motivates our question without answering it. Dynamical constraints on neural population activity shows that motor-cortex activity follows an intrinsic flow field that monkeys cannot override. In one challenge, success dropped from 92.9% to 50.7%. This suggests the stimulation optimum is tied to slowly changing network dynamics, so drift should be structured and learnable rather than pure noise.

Brain plasticity following MI-BCI training combined with tDCS found biological changes at 4 weeks only under active stimulation, despite equal behavioral gains in both arms. Those changes were fractional anisotropy, a diffusion-MRI index of white-matter integrity, and resting motor threshold. This tells us durability endpoints must go beyond immediate behavior. It used fixed settings throughout, so it says nothing about tuning.

Research Program

The program has three stages, each able to kill the idea cheaply before the next.

The first stage is a simulated testbed with known ground truth. We build a small neural-population model with three ingredients. Its acute response to stimulation is an inverted-U in intensity. Its optimal setting drifts slowly across "sessions". It carries a slow plasticity variable that grows only when the delivered effective dose stays above a threshold, and otherwise relaxes back to baseline. The plasticity variable is the elastic-band picture of durability made explicit. We vary drift speed and peak sharpness, and compare tuning rules on retained plasticity after stimulation stops. This stage tells us under which drift regimes the idea can help at all, and gives priors for $H$ and $\alpha$.

The second stage is retrospective analysis on real stimulation data. The primate optogenetic micro-electrocorticography dataset from Temporal Basis Function Models for Closed-Loop Neural Stimulation spans 40 sessions in two macaques, with paired light pulses delivered in S1 and M1. Treating sessions as time, we fit a response model per session over whatever stimulation variables the recordings span, such as pulse timing relative to ongoing state. We then ask two questions. How far does the best setting move between sessions? Does local sharpness at session $k$ predict the loss in steering performance when that setting is reused at session $k+j$? We also use the 85-subject motor-imagery EEG dataset from Unsupervised Adaptive Deep Learning Method for BCI Motor Imagery Decoding to estimate realistic rates of cross-session drift in human brain signals. We measure drift as the shift in channel covariance that its Euclidean Alignment step corrects. That estimate calibrates the time kernel's prior.

A risk we name up front: the optogenetic dataset may not vary stimulation parameters densely enough to map a surface. If so, this stage reduces to measuring cross-session transfer of a single-session model. That is weaker, but still informative about drift.

The third stage is a prospective home-based trial following the design of Personalized home based neurostimulation via AI optimization augments sustained attention, extended to a multi-week course with follow-up after treatment ends. Arms receive transcranial random noise stimulation within the same 0.1–1.6 mA envelope over roughly 4 weeks of sessions, with the same tablet-based sustained-attention task. Assessments occur at the end of treatment and at 2 and 6 weeks after cessation.

Baselines

The naive reference points are sham and a fixed 1.5 mA dose. In the source trial, fixed 1.5 mA did not separate from sham. Both run with identical session schedules, blinding and outcome tests, so any arm must beat them to matter at all.

The strongest existing method is personalized Bayesian optimization as described in Personalized home based neurostimulation via AI optimization augments sustained attention. It finds the setting with the best predicted acute response from baseline sessions and then holds it. The paper reports a significant advantage over sham and fixed dosing for low-baseline performers (p = 0.003). The file gives no effect size, and there is no reported result found on Lacuna for multi-week retention under personalized Bayesian optimization.

We also run a stronger variant of it that re-optimizes every session with the same exploration budget as our method. This is the fairest version of "just keep tuning". All tuning arms share the Gaussian-process implementation, the input features (baseline performance, head circumference), the number of exploratory settings per session, and the intensity envelope.

Beating these baselines means two things. First, end-of-treatment performance statistically non-inferior to peak tuning. Second, a significantly smaller post-cessation decline, with larger area under the performance curve from the end of treatment to 6-week follow-up.

Two ablations isolate the key ingredient. The first uses the time-aware Gaussian process but a peak acquisition ($H=0$). The second uses the future-quantile acquisition but a static kernel. If the gain comes only from modeling time, and not from preferring flat regions, the first ablation will match the full method.

In the retrospective primate analysis, the reference is the within-session steering area under the curve of about 0.70 reported for the temporal basis function model. We compare cross-session transfer of the peak-chosen versus robust-chosen settings against that within-session ceiling. In simulation, the reference is an oracle that knows the true drifting optimum.

Why not just…?

Why not just re-run Bayesian optimization at every session? This is the most serious threat. If drift is slow relative to the tuning budget, frequent re-tuning may track the optimum well enough. Its cost is that every session spends exploratory trials at deliberately uncertain settings, which lowers the effective dose delivered. The every-session re-tuning baseline, run at matched exploration budget, settles this directly. If it matches our method on post-cessation retention, we will report that the problem is tracking, not robustness, and drop the robustness claim.

Why not just absorb drift with unsupervised alignment, as in motor-imagery decoders? Euclidean Alignment plus online batch-normalization updates raised online accuracy from 77.5% to 88.3% on the 85-subject dataset in Unsupervised Adaptive Deep Learning Method for BCI Motor Imagery Decoding. But it re-normalizes the input to a decoder. It does not tell you which current to deliver, because the response to stimulation, not the signal statistics, is what drifts. We still use it as an input-normalization step in the primate analysis, so any benefit it provides is shared by all arms.

Why not just shrink the personalized dose toward the population average, or pick a conservatively lower dose? Shrinkage is a cheap robustness heuristic and could capture much of the gain if flat regions tend to lie below peaks. It fails in principle when a person's flat region lies on the other side of their peak. It also ignores each person's measured drift. We add a shrinkage arm to the simulation and retrospective stages. It enters the prospective trial only if it matches our method there.

Why not just escalate dose, as in pulse-count trials? The protocol in Different doses of intermittent theta burst stimulation compares 600, 1,200 and 1,800 pulses under an implicit monotonic dose–response assumption. The inverted-U found in the attention trial means that more current past the individual optimum hurts rather than helps. Escalation therefore addresses cumulative dose while ignoring where on the surface each session lands.

Decisive Checks

The idea lives or dies on a few observations. In the retrospective primate analysis, local sharpness at the chosen setting must predict cross-session performance loss. A null correlation there means the proposed mechanism is absent, and we stop before the human trial.

In simulation, the robust rule must beat peak tuning on retained plasticity across a meaningful range of drift speeds. It must also lose, as expected, when drift is zero; that loss is a sanity check.

In the prospective trial, the full method must be non-inferior to peak tuning at the end of treatment and show significantly less decline by 6 weeks. The time-kernel-only ablation must fall short of the full method. We will also report whether the time-aware model retains participants who would have been excluded for volatile baselines. The source trial lost 27% of participants that way, and retaining them would be an independent practical benefit.

A positive result would give a concrete, testable account of one cause of washout: tuning for today's peak. It would also give a rule that clinical tuning systems can adopt without new hardware.