People with paralysis can feel touch again through an electrode array placed in the touch area of the brain, which delivers small electrical pulses known as intracortical microstimulation (ICMS). One finding makes this harder than it sounds. Changing the pulse frequency does not simply make sensations stronger. On some electrodes higher frequency feels weaker, on others it feels stronger, and the quality of the sensation shifts between "tapping," "pressure," "buzzing," and "tingle." Today each electrode in each new patient has to be mapped by hand, through long sessions where the patient rates many stimuli. The field also has only two or three such patients per study, so no one knows which parts of a map carry over to the next patient and which are personal quirks.
We propose a calibration method that reuses what the few existing patients have in common while treating each new patient as a case that may break the pattern. It rests on two regularities. Electrodes fall into a small number of frequency-response types, and electrodes of the same type sit next to each other in millimeter-scale patches. A new patient's map can therefore be inferred from a sparse, well-chosen set of measurements instead of a full sweep. The method also reports, electrode by electrode and patient by patient, when its borrowed assumptions are failing, so it can fall back to exhaustive mapping in those places. That second property is how we bring the methodology of single case studies into neuroprosthetic engineering.
Core Claim
The evidence on cortical frequency tuning comes from very few people, but it has a structure that can be learned. Perception of microstimulation frequency in human somatosensory cortex stimulated Utah arrays in two participants with spinal cord injury. It found that frequency changes perceived intensity in a non-monotonic, electrode-dependent way, and that electrodes cluster into three types: low-frequency-preferring, intermediate-frequency-preferring, and high-frequency-preferring. Perceptual responses to microstimulation frequency are spatially organized in human somatosensory cortex showed that these types form spatial patches of roughly 1–2.5 mm and that each type goes with a characteristic sensation quality. Both papers motivate our approach. Neither builds a predictive model, and neither tests whether the three curve shapes found in two people hold in a third.
Our claim is that a hierarchical model can cut the number of psychophysical trials needed to label every electrode's frequency type substantially, at matched labeling accuracy, without losing accuracy on participants who were not used to fit it. The model's top layer holds three response-curve shapes shared across participants. Its participant layer holds that person's patch layout, encoded as a spatial smoothness prior over the electrode grid. The claim is falsifiable on the existing low-N data. The model's posterior-predictive checks are what make it a scientific instrument rather than a shortcut: they flag participants whose responses the shared shapes cannot explain, which is the "idiosyncratic outlier versus general pattern" question that clinical case reports raise but cannot answer.
The Model
Let $I_{e}(f)$ be the perceived intensity, measured by magnitude estimation, when electrode $e$ is stimulated at frequency $f$ with amplitude and train duration held fixed. We write
$$ I_{e}(f) = a_e \, g_{c_e}(f) + \epsilon, $$
where $c_e \in \{\text{LFP}, \text{IFP}, \text{HFP}\}$ is the electrode's type, $g_c$ is a smooth response curve shared across all participants, $a_e$ is an electrode-specific gain that absorbs threshold differences, and $\epsilon$ is rating noise with a participant-specific variance. The type labels $c_e$ get a Potts prior over the array grid. Under this prior, neighboring electrodes are rewarded for sharing a label, and a coupling strength $\beta$ is learned per participant. The model therefore expects patches but does not assume their size or location. The curves $g_c$ are fit on the participants available as training cases.
During calibration of a new participant, the next trial is the (electrode, frequency) pair whose answer most reduces uncertainty about the map's labels. Measuring one electrode partly labels its neighbors, so the method can skip electrodes deep inside a confidently labeled patch and spend trials on patch borders.
This structure bears on one of the field's open problems. Larger amplitudes spread current into neighboring populations that may have different frequency preferences. If a border is labeled, the encoder can avoid stimulating electrodes on that border with high-amplitude frequency codes. We report this as a secondary analysis, not a headline claim.
Research Program
The work has three parts.
First, we fit the hierarchical model to the published frequency-sweep data from the participants in Perception of microstimulation frequency in human somatosensory cortex. This requires collaborating with the originating group, because no ICMS frequency dataset is listed as a Lacuna resource. We also fit it to any participants added since. The participants in Evoking stable and precise tactile sensations via multi-electrode intracortical microstimulation of the somatosensory cortex, who have been followed for 2–7 years, also give a test of whether labels stay stable over time.
Second, before any live use, we run leave-one-participant-out replay. We hold out one participant, fit the shared curves on the others, and then simulate calibration on the held-out participant by drawing trials from that person's recorded responses. With N of 2–4 this is a weak test, and we say so. It is still the only honest test of transfer available, and each held-out participant is reported as a separate case rather than averaged away.
Third, because invasive cohorts will stay tiny, we validate the transfer logic where N can be larger. Electrical nerve stimulation for sensory-neural pathway reconstruction in upper-limb amputees showed in four amputees that non-invasive surface stimulation of the residual limb produces frequency-dependent sensation types (5 Hz pat, 50 Hz vibration, 200 Hz pressure) at stable sites. It also showed that map quality depended on how the limb was injured, with mechanical trauma giving a detailed map and burn trauma a poor one. Surface stimulation is non-invasive, so we can recruit a cohort of around 20. That is enough to ask a question no invasive study can: how many training participants does the shared-curve layer need before transfer saturates, and do the outlier checks correctly flag people with poor maps, such as burn-injured limbs? The peripheral and cortical settings differ in physiology. What we transfer to the cortex is the method and its failure diagnostics, not the parameters.
Baselines
The naive reference point is the exhaustive per-electrode frequency sweep used in Perception of microstimulation frequency in human somatosensory cortex: every electrode is tested across the full 20–300 Hz range with repeated magnitude estimates and then clustered. Its labels serve as ground truth, and its trial count serves as the cost ceiling. I found no reported result on Lacuna for the number of trials or session-hours this sweep requires. We will measure that cost directly from the replayed data instead of quoting it.
The strongest simple competitor is a coarse sweep plus spatial smoothing. Each electrode gets only three frequencies, roughly one per type's peak region, and is labeled by its argmax. The labels are then smoothed with a majority vote over grid neighbors. There are only three types and they cluster, so this method uses the same two regularities as ours without any shared curve model or adaptive trial selection. It will get exactly the same replayed data, the same total trial budget, and its frequency triplet tuned on the training participants just as our curves are. We consider our method to beat it only if it reaches the same label agreement with ground truth, measured by Cohen's kappa against the full-sweep labels, using clearly fewer trials on held-out participants. This must hold on each held-out participant, not merely on average.
Two ablations isolate the key ingredients. The first removes the Potts spatial prior, setting $\beta = 0$, which leaves per-electrode Bayesian inference with shared curves. The second keeps the spatial prior but selects trials at random. If the spatial prior delivers all the gains, the adaptive selection is unnecessary and we will drop it.
For downstream benefit, the reference numbers come from the multi-electrode work. Biomimetic multi-channel microstimulation of somatosensory cortex conveys high resolution force feedback for bionic hands reports 92.5% compliance-discrimination accuracy with multi-channel biomimetic feedback versus 75% with single-channel linear feedback. Evoking stable and precise tactile sensations via multi-electrode intracortical microstimulation of the somatosensory cortex reports localization error falling from 42% with single electrodes to 7% with quartets. Our aim on these tasks is non-inferiority: encoders configured from the fast calibration should not fall measurably below encoders configured from full sweeps.
Why not just…?
Why not just use the coarse three-frequency probe with neighbor smoothing? This is the real threat. With three types and clear patches, it may match the full method. The proposal is built so that this outcome is informative rather than embarrassing. Our method has two plausible advantages. It can skip electrodes entirely, while the coarse probe still touches every electrode. Its predictive checks also detect participants whose curves do not match the shared shapes, while argmax labeling silently forces every electrode into one of three bins. If the matched-budget comparison shows no trial savings and no better outlier detection, we will recommend the coarse probe and report the spatial prior as the useful finding.
Why not just ignore frequency and modulate amplitude? Restored tactile sensation improves neuroprosthetic arm control showed that linear amplitude feedback cut median ARAT trial time from 20.9 s to 10.2 s in one participant, so amplitude coding clearly helps. The multi-channel results cited above show that it has a ceiling, however. Single-channel linear feedback reaches about 8 discriminable force levels and 75% compliance accuracy, against about 19.5 levels and 92.5% for multi-channel biomimetic feedback. In Biomimetic stimulation patterns drive natural artificial touch percepts using intracortical microstimulation in humans, patterns that co-modulate amplitude and frequency were preferred in 75% of trials and used 5.5 µC of charge versus about 9 µC. Those frequency-modulating encoders need to know each electrode's frequency type, which is the quantity we calibrate.
Why not let the patient self-tune on a tablet? Conveying tactile object characteristics through customized intracortical microstimulation of the human somatosensory cortex did this with three participants and reached high satisfaction. Blinded replay accuracy was only about 37% against a 20% chance level, however, and it declined over time. Exploration styles also varied widely, so the resulting maps mix neural tuning with search strategy. Self-tuning answers "what does this person like," while our method answers "what is this electrode's tuning." The two are complementary, and our labels could restrict a tablet search to electrodes and frequencies that actually produce the intended sensation quality.
Why not just wait for larger cohorts? Every human ICMS study in the file has between one and four participants, because implantation requires neurosurgery. Waiting does not fix the calibration burden on each new patient. The surface-stimulation cohort is our partial substitute for a large invasive cohort, and it tests the transfer logic itself.
Decisive Checks
The first decisive check is leave-one-participant-out replay on the ICMS frequency data. The outcome is the number of trials needed to reach a fixed kappa against full-sweep labels, compared across our method, the coarse-probe baseline, and the two ablations, and reported separately for each held-out participant. The second check repeats this on the surface-stimulation cohort, with a learning curve over the number of training participants and a test of whether the predictive checks flag participants with poor maps, as in the burn-injury case reported in Electrical nerve stimulation for sensory-neural pathway reconstruction in upper-limb amputees. The third check is prospective. In the next available implanted participant, we compare biomimetic encoders configured from our calibration and from a full sweep on compliance discrimination and localization, using the 92.5% and 7% figures above as targets for non-inferiority.
A clean negative result would also be useful. That would happen if the shared curves fit poorly on held-out participants, or if the outlier checks fire on most new participants. Either outcome would support the skeptical view that cortical frequency maps are mainly idiosyncratic and that per-patient exhaustive mapping cannot be avoided. That view is currently asserted from N ≤ 3 and has never been tested.
References
- Perception of microstimulation frequency in human somatosensory cortex
- Perceptual responses to microstimulation frequency are spatially organized in human somatosensory cortex
- Evoking stable and precise tactile sensations via multi-electrode intracortical microstimulation of the somatosensory cortex
- Electrical nerve stimulation for sensory-neural pathway reconstruction in upper-limb amputees
- Biomimetic multi-channel microstimulation of somatosensory cortex conveys high resolution force feedback for bionic hands
- Restored tactile sensation improves neuroprosthetic arm control
- Biomimetic stimulation patterns drive natural artificial touch percepts using intracortical microstimulation in humans
- Conveying tactile object characteristics through customized intracortical microstimulation of the human somatosensory cortex