When a cortical neuron's cell body is held slightly depolarized, the action potentials traveling down its axon get wider — by about 12.6% under physiological conditions — because the depolarization creeps electrotonically into the axon and partially inactivates a class of slow potassium channels (Kv1) that normally reset the spike (Somatic Membrane Potential and Kv1 Channels Control Spike Repolarization in Cortical Axon Collaterals and Presynaptic Boutons; Selective control of cortical axonal spikes by a slowly inactivating K+ current). Wider spikes hold presynaptic calcium channels open longer, boosting release probability and, in a formal synaptic channel model, cutting the receiver's detection error from about $10^{-1}$ to $10^{-2}$ at a single synapse and below $10^{-4}$ across four terminals — all at fixed firing rate (P/Q and N Channels Control Baseline and Spike-Triggered Calcium Levels in Neocortical Axons and Synaptic Boutons; Synaptic Channel Model Including Effects of Spike Width Variation). Spike width is thus a genuine analog gain dial on synaptic transmission, and clinical stimulation protocols that chronically polarize neurons (tDCS over ten sessions, rTMS over two weeks) plausibly turn it.
Here is the problem: every measurement of this dial comes from acute preparations lasting seconds to hours, while every clinical stimulation study that could chronically engage it stops measuring at or before two weeks. No reported result found on Lacuna for any longitudinal measurement of axonal spike width beyond six hours — the longest anchor in the corpus is the 6-hour rat phrenic-nerve stimulation model, which tracked compound muscle action potentials but never individual spike repolarization kinetics (A preclinical rat model for bilateral phrenic nerve stimulation during mechanical ventilation). Meanwhile clinical neurostimulation trials routinely show acute gains with no follow-up at all (Multisession tDCS combined with intrastimulation training improves emotion recognition in adolescents with autism spectrum disorder) or only two weeks of it (Repetitive transcranial magnetic stimulation promotes the recovery of upper limb motor dysfunction in ischemic stroke patients: a DTI-based glymphatic system imaging prospective study). This proposal asks the durability question directly: if you depolarize a neuron intermittently for weeks, does the 12.6% broadening persist, decay, or reverse — and can we estimate that answer from inevitably truncated experiments without fooling ourselves?
Core Claim
I claim two things must be built together, because each is useless without the other. First, an adaptation-aware biophysical model of the axon: a multi-compartment simulation with Kv1 slow inactivation, P/Q- and N-type calcium entry, and — the missing ingredient in every prior model — homeostatic regulation of axon initial segment potassium conductance and calcium-dependent down-regulation, so that chronic stimulation can drive the system toward its own compensatory fixed point. Second, a censoring-aware durability estimator: when chronic recordings (simulated or real) end early because the preparation degrades, that dropout is not random — stimulation stress that broadens spikes also loads boutons with calcium and accelerates seal failure, so naive survival estimators will systematically overestimate stability. The survival-analysis literature has already shown that explicitly modeling informative censoring beats treating it as random, with the SALaD method winning 109, tying 28, and losing only 19 of 156 comparisons across ten datasets (Explicitly Modeling Censoring Produces Superior Survival Predictors) — but that method has never touched an electrophysiological waveform. The combined deliverable is a durability curve for spike broadening: an estimate, with calibrated uncertainty, of the retention
$$R(t) = \Pr\!\left[W(t) - W_0 > \tfrac{1}{2}\,(W_1 - W_0) \mid \text{protocol}\right]$$
where $W(t)$ is axonal spike width at chronic-stimulation day $t$, $W_0$ the pre-stimulation baseline, and $W_1$ the day-1 broadened width; $R(t)$ is the probability the analog gain survives past half its initial boost at day $t$. The headline question is whether $R(28)$ is above or below 0.5 for realistic clinical duty cycles, and whether any truncated-experiment estimator that ignores adaptation or censoring gets that call right.
The tension that makes this nontrivial is already documented on both sides. Acute data show sustained depolarization enhances transmission through width and calcium (P/Q and N Channels Control Baseline and Spike-Triggered Calcium Levels in Neocortical Axons and Synaptic Boutons), while biophysical simulations of calcium-activated potassium feedback and repolarization currents show prolonged calcium accumulation pulls excitability back down (Simulations Suggest Information Processing Roles for the Diverse Currents in Hippocampal Neurons). Separately, depolarizing pyramidal cells recruits disynaptic inhibition that narrows the integration window for their own inputs (Membrane Potential-Dependent Modulation of Recurrent Inhibition in Rat Neocortex) — so at the circuit level, excitatory broadening and recruited inhibitory clamping compete on a timescale nobody has measured. Which force wins over weeks is exactly the open question.
Research Program
The program has three stages, each anchored to a specific empirical dataset from the corpus.
Stage one builds and calibrates the acute biophysical model. I will construct a multi-compartment cortical axon model (initial segment, nodes, collaterals, boutons) in a NEURON-style environment, with channel kinetics drawn from the standardized Hodgkin-Huxley omnimodel library, which currently fits 72% of tested ion channel models (2229 of 3083) (An ion channel omnimodel for standardized biophysical neuron modelling). Calibration targets are hard numbers, not vibes: reproducing the average 12.6% collateral spike broadening for a −70 mV to −50 mV somatic shift and the 250–300% broadening under α-dendrotoxin or 4-aminopyridine block (Somatic Membrane Potential and Kv1 Channels Control Spike Repolarization in Cortical Axon Collaterals and Presynaptic Boutons); reproducing the axonal spike lengthening from roughly 1.2 ms to 3.3 ms under potassium current block with a much smaller somatic effect (Selective control of cortical axonal spikes by a slowly inactivating K+ current); and sitting on the universal initial-segment relationship between repolarization rate ratio $y$ and spike duration $x$, namely $y = 0.12 + 0.16\,\exp((0.5 - x)/0.2)$, which holds despite a threefold difference in potassium conductance density between fast-spiking interneurons ($495.7 \pm 108$ pS/µm²) and regular-spiking pyramidal cells ($185.8 \pm 19$ pS/µm²) (Axon initial segment potassium channel density in cortical neurons). Calcium dynamics are calibrated to the two-photon dataset of 380 locations across 80 layer-5 pyramidal neurons, including the roughly 20% baseline calcium rise with a ~1 s time constant under somatic depolarization (P/Q and N Channels Control Baseline and Spike-Triggered Calcium Levels in Neocortical Axons and Synaptic Boutons).
Stage two adds the slow variables that acute experiments cannot see and runs the chronic protocols. Three adaptation mechanisms are added to the calibrated model: activity-dependent up-regulation of initial-segment potassium conductance (homeostatic plasticity that would move a pyramidal cell's $g_K$ from its $185.8$ pS/µm² baseline toward the interneuron-like regime that resists broadening — the direction and magnitude are fitted as free parameters, constrained by the two conductance endpoints already measured); calcium-dependent down-regulation of release machinery, motivated by the negative-feedback simulations cited above (Simulations Suggest Information Processing Roles for the Diverse Currents in Hippocampal Neurons); and the recruited disynaptic inhibition with its measured >5 mV depolarization threshold and 400–800 µm electrotonic space constant (Membrane Potential-Dependent Modulation of Recurrent Inhibition in Rat Neocortex). Because a full multi-compartment model over simulated weeks is expensive, I will use the event-driven multi-compartment leaky integrate-and-fire reduction — a sum of six exponential functions shown to replicate NEURON mitral-cell responses with negligible error — for the long-horizon sweeps (Multicompartment leaky integrate and fire neuron modeling with multiexponentials), reserving the full model for verification at sampled checkpoints. Chronic protocols mirror the clinical arms in the corpus: ten sessions of intermittent polarization as in the adolescent tDCS trial, and two weeks of daily stimulation as in the rTMS stroke cohort, extended virtually to 28 days.
Stage three is the estimator and the honest handling of truncation. I will generate cohorts of a few hundred simulated "preparations," each with an individual parameter draw (cell type, conductance, adaptation rate), each run under a follow-up limit drawn from the empirical anchors — 6 hours, 2 weeks, or truncated early by a calcium-stress-driven failure process that couples dropout probability to accumulated calcium load. This is precisely the setting where the censoring literature predicts naive estimators collapse: SALaD showed that decomposing inputs into event-specific, censoring-specific, and shared latent factors with maximum-mean-discrepancy invariance constraints recovers signal that non-informative-censoring models discard (Explicitly Modeling Censoring Produces Superior Survival Predictors). I will adapt that decomposition to time-series waveform features (spike width, repolarization slope, baseline calcium proxy, stimulation dose) rather than tabular clinical covariates, and validate against the one ground truth survival analysis never has in the clinic: the uncensored full-horizon simulation. The decisive downstream check uses the synaptic channel model as the functional readout — translating each day's width trajectory into detection error per its published sweep from $10^{-1}$ to below $10^{-4}$ across terminals (Synaptic Channel Model Including Effects of Spike Width Variation) — and asks whether the day-1 error floor survives to day 28. Finally, an empirical arm closes the loop: an organotypic cortical culture chronically depolarized on a duty cycle, with repeated voltage-sensitive-dye imaging of axonal spike width at days 1, 7, 14, and 28, providing the first direct durability measurement the field lacks.
Baselines
The first reference point is naive persistence: simply assume the acute 12.6% broadening measured in slices (Somatic Membrane Potential and Kv1 Channels Control Spike Repolarization in Cortical Axon Collaterals and Presynaptic Boutons) holds indefinitely under chronic stimulation, so $R(t) = 1$ for all $t$. This is the implicit assumption behind interpreting acute synaptic-enhancement effects as therapeutic mechanisms, and beating it means showing the model's predicted retention curve deviates from flat with mechanisms that are individually validated. It costs no compute; the comparison is whether the culture arm's measured widths at day 28 sit on the flat line or the model's curve.
The strongest existing competitor for the estimation half is SALaD itself, applied off-the-shelf to the simulated chronic cohorts (Explicitly Modeling Censoring Produces Superior Survival Predictors). It is the strongest because it already wins 109 of 156 comparisons against methods that ignore informative censoring, and it is domain-agnostic enough to accept engineered waveform features as inputs. To be fair I will give it identical data, identical feature budget, and identical hyperparameter-search compute as my biophysics-informed variant; beating it means materially lower root-mean-squared error on the true $R(28)$ and on the median durability half-life across simulation replicates. My variant's edge comes only from its inductive bias — using the mechanistic model to generate features (calcium load, Kv1 inactivation state) that the pure statistical method cannot observe — so if SALaD matches accuracy without them, the mechanistic half of the proposal is not earning its keep on the estimation side, and I would report that.
The strongest competitor for the modeling half is a rate-only plasticity model: a point-neuron or multi-compartment integrate-and-fire network with standard spike-timing-dependent plasticity and no spike-width variable at all, run under the identical chronic protocols and evaluated on the same downstream detection-error readout. It is the strongest because it is what the field would default to for explaining long-term changes after neurostimulation. It will be given the same compute budget and tuned with the same search over learning rates and plasticity windows. Beating it means the width-dependent pathway predicts functional outcomes (culture-arm width changes, detection error shifts) that the rate-only model cannot reproduce when firing rates are matched between conditions — the rate-matched comparison is the fair form.
Two ablations isolate the proposal's key ingredients. The adaptation-free model is the Stage-one calibration alone, with Kv1 inactivation and calcium entry but no homeostatic regulation, run over the same 28 days; it isolates whether the slow regulatory variables (not just cumulative slow inactivation) drive any predicted decay, and if it produces the same durability curve as the full model, the adaptation machinery should be dropped. The censoring-agnostic ablation is a standard Kaplan-Meier or Cox estimate on the identical truncated cohorts, treating dropout as random; it isolates whether coupling between stimulation stress and termination actually matters at the follow-up lengths in play, and if the bias it introduces is small at 2-week truncation but large at 6-hour truncation, that itself is a publishable, quantified answer about which experimental designs are safe.
Why not just…?
Why not just assume the acute effect persists, since the clinical trials showed gains anyway? Because the clinical gains themselves are exactly the unverified durability in question: the adolescent tDCS trial reported no post-session follow-up at all (Multisession tDCS combined with intrastimulation training improves emotion recognition in adolescents with autism spectrum disorder), the rTMS stroke study stopped at 2 weeks (Repetitive transcranial magnetic stimulation promotes the recovery of upper limb motor dysfunction in ischemic stroke patients: a DTI-based glymphatic system imaging prospective study), and the corpus is full of surrogate endpoints that failed to predict long-term outcomes in other domains. More concretely, the acute calcium boost that width provides has a documented negative-feedback counterweight in calcium-activated potassium conductances (Simulations Suggest Information Processing Roles for the Diverse Currents in Hippocampal Neurons), and no experiment in the corpus rules out that this feedback wins over weeks. The culture arm exists precisely to test the persistence assumption directly at day 28.
Why not just use firing-rate models with standard plasticity? Because the phenomenon in question is defined at fixed rate. The synaptic channel model shows spike broadening alone, holding rate constant, moves detection error by orders of magnitude (Synaptic Channel Model Including Effects of Spike Width Variation), and the paired-recording work shows subthreshold shifts of more than 5 mV alter inhibitory postsynaptic amplitudes and onset latencies independently of spike rate (Membrane Potential-Dependent Modulation of Recurrent Inhibition in Rat Neocortex). A rate-only model has no variable in which these effects live. That said, this is a real threat in one sense: if chronic stimulation mostly changes firing rates and the width pathway turns out to be a second-order contributor to the functional readout, the rate-only baseline will match the culture data and the width story shrinks to a biophysical footnote. The rate-matched comparison in the experiments is designed to settle exactly this, and I will report it either way.
Why not just run standard survival analysis with random censoring on whatever chronic data we get? Because the censoring literature has already quantified the cost of that assumption: the SALaD paper shows, theoretically and across ten datasets, that ignoring informative censoring discards critical signal and loses 109-of-156-style comparisons (Explicitly Modeling Censoring Produces Superior Survival Predictors). In electrophysiology the informative structure is mechanistic, not just statistical: the same stimulation stress that broadens spikes drives the calcium load that kills preparations early. The censoring-agnostic ablation will measure how big this bias actually is at realistic truncation lengths, and if it is negligible, that ablation result itself simplifies the field's practice — but the burden is on the ablation to show it, not on the assumption to be trusted silently.
Why not just use continuous open-loop stimulation and skip modeling durability at all? Because the control literature in the corpus says fixed protocols fight strong unmodeled variability: 97.4% of recorded channels in the primate µECoG dataset exhibited state-dependent responses to stimulation (Temporal Basis Function Models for Closed-Loop Neural Stimulation), and continuous deep brain stimulation was shown to waste 20%–55% more energy than adaptive closed-loop control at equal efficacy (Offline Learning of Closed-Loop Deep Brain Stimulation Controllers for Parkinson Disease Treatment). Both of those controllers, however, reason over 20 ms windows or single-session beta power — timescales far too short to see Kv1 slow inactivation on the order of seconds, let alone weeks of homeostatic drift. The durability curve this proposal produces is exactly the missing slow variable that a future closed-loop controller would need, and I will demonstrate this by showing that a controller conditioned on the model's predicted channel-state trajectory maintains target spike width at day 28 where a short-window controller, retrained daily on acute observations, drifts.
Decisive Checks
The proposal succeeds or fails on four concrete checks. First, calibration fidelity: the Stage-one model must reproduce the cited numbers — 12.6% and 250–300% broadening (Somatic Membrane Potential and Kv1 Channels Control Spike Repolarization in Cortical Axon Collaterals and Presynaptic Boutons), the 1.2 ms to 3.3 ms block-induced lengthening (Selective control of cortical axonal spikes by a slowly inactivating K+ current), and the universal $dV/dt$-duration relationship (Axon initial segment potassium channel density in cortical neurons) — within their reported spread before any chronic extrapolation is trusted. Second, estimator bias: against uncensored ground truth, the censoring-aware estimate of $R(28)$ must beat both the naive-persistence reference and random-censoring survival estimates by a margin that holds across simulated truncation regimes anchored to the 6-hour and 2-week empirical ceilings (A preclinical rat model for bilateral phrenic nerve stimulation during mechanical ventilation; Repetitive transcranial magnetic stimulation promotes the recovery of upper limb motor dysfunction in ischemic stroke patients: a DTI-based glymphatic system imaging prospective study). Third, the rate-matched test: with firing rates equated, the width-dependent pathway must predict functional changes the rate-only model cannot. Fourth, and finally decisive, the culture arm: the first repeated-measures axonal spike-width trajectory past 24 hours, landing on the predicted retention curve rather than the flat persistence line — or refuting both, which would send the model back to the drawing board and still be the first datum the field has ever had at that horizon.
References
- Somatic Membrane Potential and Kv1 Channels Control Spike Repolarization in Cortical Axon Collaterals and Presynaptic Boutons
- Selective control of cortical axonal spikes by a slowly inactivating K+current
- P/Q and N Channels Control Baseline and Spike-Triggered Calcium Levels in Neocortical Axons and Synaptic Boutons
- Synaptic Channel Model Including Effects of Spike Width Variation
- A preclinical rat model for bilateral phrenic nerve stimulation during mechanical ventilation
- Multisession tDCS combined with intrastimulation training improves emotion recognition in adolescents with autism spectrum disorder
- Repetitive transcranial magnetic stimulation promotes the recovery of upper limb motor dysfunction in ischemic stroke patients: a DTI-based glymphatic system imaging prospective study
- Explicitly Modeling Censoring Produces Superior Survival Predictors
- Simulations Suggest Information Processing Roles for the Diverse Currents in Hippocampal Neurons
- Membrane Potential-Dependent Modulation of Recurrent Inhibition in Rat Neocortex
- An ion channel omnimodel for standardized biophysical neuron modelling
- Axon initial segment potassium channel density in cortical neurons
- Multicompartment leaky integrate and fire neuron modeling with multiexponentials
- Temporal Basis Function Models for Closed-Loop Neural Stimulation
- Offline Learning of Closed-Loop Deep Brain Stimulation Controllers for Parkinson Disease Treatment