A brain-computer interface translates spikes recorded by an array of microelectrodes into movement commands. These arrays are made of rigid silicon or metal sitting in brain tissue that is about a million times softer. Every heartbeat and head movement makes the tissue shear against the electrodes. Over weeks, electrodes crack, scar tissue grows around them, and neurons near them die. The damage is not random: it concentrates on the outer edge of the array. Decoders trained on one day therefore degrade on later days. Yet every widely used decoder still outputs a single number with no indication of whether it can be trusted today. We propose to give a single trained decoder a cheap, calibrated "I might be wrong" signal aimed at this failure mode.
The idea is simple. Some recent decoders can accept any subset of channels. We run such a decoder many times at once on different random subsets of the electrodes and measure how much the predictions disagree. When a few electrodes have drifted or broken, subsets that include them disagree with subsets that exclude them, so the spread rises exactly when the sensors are going bad. We then convert that spread into prediction intervals with a guaranteed error rate using only the few labeled seconds of a new session's calibration data. This matters because the realistic setting is a single patient or monkey, a short recalibration run, and hardware that cannot be swapped out.
Core Claim
For decoders that treat the electrodes as an unordered set, disagreement across channel subsets tracks cross-session decoding error better than the usual uncertainty methods. Those methods either perturb network weights or train several networks: Monte Carlo dropout keeps random dropout active at test time, and deep ensembles train several independently initialized networks. Weight-based spread measures the model's ignorance about its own parameters. Channel-subset spread measures how much the answer depends on individual sensors, and sensor corruption is the dominant reason these decoders fail. After a light rescaling on the few-shot calibration split, this signal yields intervals whose coverage holds on held-out sessions. It also supports "abstain or recalibrate" decisions that preserve accuracy on the retained data.
Mechanical mismatch plays two roles in this proposal. First, it explains why sensor-level corruption, not a change in the underlying neural computation, is the expected shift. Second, it is a stress test. We check whether the signal responds to realistic, edge-concentrated degradation and not just to random deletions.
What the Prior Work Establishes, and What It Leaves Open
Three groups of papers show that degradation is spatially structured. Forrest et al. simulated micromotion strain in Utah arrays with a finite element model, a numerical mechanics simulation. They found perimeter and corner electrodes are strained far more than interior ones (p < 0.0001), and that strain correlates with one-year signal quality in human motor cortex (SNR ρ = −0.23). Patel et al. found a 63% neuron loss near the electrode shanks in one macaque. Insulation cracks (40.3% of shanks), metal cracks (39.7%) and tip breakage (22.3%) concentrated in the outer rows. Urdaneta et al. showed that damage in rat cortex also varies with depth. All three are descriptive studies on two to six animals, and none attaches a decoder or a reliability estimate. They motivate the proposal but cannot be the training data.
Forrest et al. also report a complication that constrains our design. Higher strain went with lower 1 kHz impedance (ρ = −0.39 in human motor cortex, −0.41 in macaque V4). They attribute this to cracked insulation or fluid creating leakage paths. Any reliability signal built on impedance thresholds would therefore mark some of the most damaged electrodes as healthy. This is one reason we derive uncertainty from the decoder's behavior and not from hardware diagnostics.
The decoding side supplies the method and the benchmark. SPINT is a transformer that reads electrodes as an unordered set. It is trained by randomly deleting 0–100% of channels, reaches R² ≈ 0.52 on FALCON M1 with 80% of channels removed, and runs at about 0.13 of the real-time budget. Its authors use channel dropout only as training augmentation and never read uncertainty out of it. That missing step is what we add. FALCON measures accuracy under cross-session drift with small calibration splits. On M1 the retrained-daily "oracle" decoder reaches R² ≈ 0.75 and day-zero static decoders fall well short. FALCON has no uncertainty track. NDT3 and seegnificant attack small cohorts through pretraining and electrode-position embeddings, and both output only point predictions.
The uncertainty literature both warns and guides. Bench et al. benchmarked eight methods on pulse-sensor signals. They found no universal winner and a sharp loss of local calibration when test subjects were unseen. This is why we evaluate local metrics and held-out sessions, not only a global calibration score. Tveter et al. showed on scalp EEG that discarding the most uncertain windows cuts error by up to 32%. We borrow their selective-rejection protocol and move it to continuous intracortical decoding. Xia et al. show that single-pass evidential learning needs class-balanced priors to stay calibrated in small cohorts, and they formulate it only for classification. Evidential regression is therefore a baseline we must build ourselves, not one we can take off the shelf.
Research Program
The base model is a SPINT-style set decoder trained exactly as in its paper. Nothing is added at training time, so any gain comes from how the model is read out, not from a different model.
At test time, for each time bin we draw $K$ channel subsets with a fixed keep fraction and run them as a single batch. The batched pass keeps latency close to one forward pass given SPINT's headroom. The prediction is the mean over subsets. The raw uncertainty $u(x)$ is the per-dimension standard deviation across subsets.
On each new session's few-shot labeled calibration split, we apply split conformal prediction, a distribution-free way to turn any uncertainty score into intervals with a stated error rate. The procedure uses the normalized score $s_i = |y_i - \hat y_i| / u(x_i)$ and sets the interval half-width to $\hat q \cdot u(x)$. Here $y_i$ is the true output, $\hat y_i$ the prediction, and $\hat q$ the $\lceil (n+1)(1-\alpha) \rceil / n$ empirical quantile of the $n$ calibration scores, for target miscoverage $\alpha$. Neighbouring time bins are correlated, so scores are computed over contiguous blocks, not bins. We report how coverage behaves as $n$ shrinks, since small calibration sets are the regime of interest.
We then study three questions. The first is whether the signal is sensor-specific: does it rise more under channel corruption than under changes in behavior? The second is whether it responds to structured degradation. Here we compare subsets drawn uniformly with subsets that deliberately exclude the perimeter when electrode positions are known. The third is whether it supports decisions: abstain on uncertain bins, or trigger recalibration when session-level uncertainty crosses a threshold.
The structured stress test applies synthetic degradation to real held-out sessions. We impose edge-first channel loss, amplitude attenuation, and a raised noise floor on perimeter channels. The last mimics cracked insulation that lowers impedance while destroying signal. This requires the channel-to-electrode map for the FALCON arrays. If that map is not released, we fall back to clustered degradation on a channel-similarity graph. We will say explicitly that the fallback is a weaker proxy for the edge effect.
Baselines
The naive reference is the SPINT point decoder with a constant-width conformal interval fitted on the same calibration split. It achieves nominal coverage but cannot tell good bins from bad ones. Any proposal must beat it on interval width at matched coverage and on error-versus-retained-fraction curves.
The strongest existing competitors are weight-based uncertainty methods applied to the same SPINT backbone. We will run a deep ensemble of five independently seeded SPINT models and Monte Carlo dropout over SPINT's weights. Both will be conformalized identically to our method. Bench et al. found ensembles and Monte Carlo dropout, combined with post-hoc recalibration, among the best methods: Monte Carlo dropout with isotonic regression reached ECE = 0.048 on DeepBeat atrial-fibrillation detection. That is a different task and modality, so it guides method selection rather than serving as a number to beat. There is no reported result found on Lacuna for uncertainty calibration, interval coverage, or selective decoding on any FALCON dataset. The comparison therefore has to be established here.
Fairness has two parts. The ensemble gets five times the training compute and we report it that way. All methods get the same number of test-time forward passes, tuning budget, and calibration split. Beating the ensemble means equal-or-better coverage with narrower intervals and a lower area under the error-versus-retained-fraction curve on held-out sessions. It must also show a larger response under structured degradation, at one-fifth the training cost. Matching the ensemble at one-fifth the cost would still be a useful result, and we would report it as such.
Three further baselines round out the comparison. A Kalman filter's built-in covariance is the classical linear reference for accuracy-plus-uncertainty. An evidential regression head on SPINT is the single-pass competitor, since Xia et al. only cover classification. An input-side drift detector scores each bin by how far per-channel firing statistics sit from the training distribution.
The key ablation holds everything fixed and swaps the source of spread. Channel-subset spread is compared with weight-dropout spread from the same trained model. This isolates whether sensor-level perturbation is really the useful ingredient. A second ablation compares uniform subsets with perimeter-aware subsets.
Why not just…?
Why not just use a deep ensemble of decoders? This is the real threat. Ensembles are the most reliable general-purpose method in Bench et al.'s comparison, and they may simply match our method. Our argument is mechanistic. Independently trained set decoders all learned to be robust to channel loss, so they may agree with each other on corrupted input. Channel subsets from one model, by contrast, directly expose which sensors the prediction hinges on. The weight-versus-channel ablation and the structured stress test settle this. If ensembles track error equally well under edge-first degradation, the contribution shrinks to "the same quality at a fifth of the training cost," and we will say so.
Why not just monitor the inputs for drift and skip decoder uncertainty? An unsupervised detector on firing statistics is cheap and needs no labels. However, it flags any change, including changes in behavior the decoder handles fine. Forrest et al.'s impedance result also shows that hardware-level health indicators can point the wrong way. We include the detector as a baseline and ask a direct question: which signal better predicts the per-session R² drop, and which better selects bins for rejection? If the input detector wins on session-level alarms but loses on bin-level selection, the two are complementary, and we would recommend combining them.
Why not just use a Kalman filter, which already gives uncertainty? Its covariance is exact only if the linear model is right. FALCON reports that day-zero static decoders fall far below the R² ≈ 0.75 oracle on M1. NDT3 reports that a transformer fine-tuned on 10% of the data beats a linear model trained on all of it. Calibrated uncertainty around a much worse point estimate is not a useful trade. We still report the Kalman filter's coverage, because it is the natural reference for whether uncertainty is calibrated across sessions.
Why not just recalibrate every session and forget uncertainty? FALCON's few-shot alignment baselines give only marginal gains, and recalibration costs user time. An uncertainty signal is what tells a clinician when recalibration is needed. We test this with a recalibrate-on-alarm policy: it should reach near-daily-recalibration accuracy while using fewer recalibration sessions.
Proposed Experiments
We will use FALCON M1 (EMG decoding, 16 channels), M2 (finger kinematics), and H1 (7-degree-of-freedom arm control). We train on held-in sessions and use each held-out session's official few-shot calibration split for conformal fitting. For H1 and M2 we reproduce baseline accuracies ourselves, since the file has no reported numbers for them.
The metrics are the following. Accuracy is measured with R². Calibration uses empirical coverage and interval width at α = 0.1 and 0.2, together with the continuous ranked probability score, a single number rewarding forecasts that are both sharp and calibrated. Local calibration uses the ENCE metric from Bench et al., which checks whether predicted error sizes match actual errors within bins of predicted uncertainty. Selective decoding is measured as R² versus retained fraction, following Tveter et al., plus the area under that curve. At the session level we report the rank correlation between mean uncertainty and R² loss relative to the oracle.
Three checks decide the question. First, on real held-out sessions, channel-subset spread should correlate more strongly with per-session R² loss than weight-based spread from the same model. Second, under edge-first synthetic degradation of increasing severity, our uncertainty should rise steadily while coverage stays near nominal. The comparison methods should show smaller increases or lose coverage. Perimeter-aware subsets should beat uniform subsets only when degradation is actually edge-structured, and not under random loss; that contrast is a built-in control. Third, conformal coverage should remain within a few points of nominal as the calibration split is cut to its smallest usable size.
If the first check fails, meaning weight-based spread ranks sessions as well as ours, the central claim is false. We would then report the benchmark and stress test as the contribution.
References
- Finite element model predicts micromotion-induced strain profiles that correlate with the functional performance of Utah arrays in humans and non-human primates
- Utah array characterization and histological analysis of a multi-year implant in non-human primate motor and sensory cortices
- Layer-dependent stability of intracortical recordings and neuronal cell loss
- SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding
- Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark
- A Generalist Intracortical Motor Decoder
- Neural decoding from stereotactic EEG: accounting for electrode variability across subjects
- A systematic evaluation of uncertainty quantification techniques in deep learning: a case study in photoplethysmography signal analysis
- Advancing EEG prediction with deep learning and uncertainty estimation
- Uncertainty-Aware Health Diagnostics via Class-Balanced Evidential Deep Learning