Long-term unsupervised recalibration of cursor-based intracortical BCI using a hidden Markov model

• Intracortical BCIs suffer from decoder drift over time, limiting real-world usability.¹
• An HMM-based unsupervised recalibration method addresses drift in cursor-based intracortical BCIs.¹
• The approach directly improves long-term, at-home usability for cursor control without supervised retraining.^{1 1}

Weekly enrichment (2026-07-20)

• The method, called PRI-T (Probabilistic Retrospective Inference of Targets), uses a hidden Markov model on decoder outputs to infer which target the user is moving toward, generating pseudo-labels plus confidence estimates that weight each label during unsupervised decoder retraining.^{2 3}
• Because PRI-T operates on cursor/decoder outputs rather than neural latent representations, it is signal-agnostic (deployable with local field potentials, ECoG, or EEG) and can retrain arbitrary multi-layer decoders, unlike latent-space alignment methods that update only early layers.²
• Development used 73 recording sessions spanning five years from participant T5 in the BrainGate2 pilot clinical trial, with neural activity recorded from two 96-channel silicon microelectrode arrays (Blackrock Neurotech) in the hand-knob of dorsal motor cortex.^{2 4}
• Nonstationarity was quantified as subspace drift in decoder weights: within-session tuning similarity was ~0.75 but dropped over 1–2 weeks of separation, and the median coefficient of determination (R²) fell from ~0.39 to ~0.21.²
• PRI-T outperformed distribution-alignment methods (FA stabilization and ADAN) in large-scale closed-loop simulations over two months and in a closed loop with a human iBCI user over one month.^{2 3}
• Across 239 offline session pairs (two-sided Wilcoxon rank-sum tests), ADAN versus PRI-T reached P=0.0015 and PRI-T versus PRI-T-plus-stabilizer reached P<0.001; distribution-matching methods accumulated compounding error and eventually diverged over long timescales.²
• The approach also performed well offline on freeform datasets of a person using a home computer with an iBCI, supporting at-home, self-supervised recalibration.^{2 3}
• Clinically, using task structure to bootstrap a noisy decoder into a high-performing one removes supervised recalibration downtime—when users cannot operate their device—which the authors frame as a major barrier to translating BCIs.²

Footnotes

1. https://news.google.com/rss/articles/CBMiX0FVX3lxTFBrazNNTE9XM3dhV1A1eW9TR095eFRsbFZtWFFnTUo4Z3RhNjVKWDdTTG1XZkxnM2xEdVgtUTh6RXRqUExFYTQtekZVUWJwNTNJLXlKY3pLWFdsZ2lfMFpr?oc=5 ↩ ↩² ↩³ ↩⁴

2. https://www.nature.com/articles/s41551-025-01536-z ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

3. https://doi.org/10.1038/s41551-025-01536-z ↩ ↩² ↩³

4. https://pmc.ncbi.nlm.nih.gov/articles/PMC9915729/ ↩