An electrocorticographic speech brain-computer interface maintained stable offline syllable decoding performance across a 25-month evaluation period in a participant with slowly progressive bulbar variant amyotrophic lateral sclerosis, despite an observed widespread decline in active spectral signal-to-noise ratio. The longitudinal study provides empirical evidence that speech-related population neural information, spatial decoding patterns, neural class structure, and task-related neural modulation can remain comparatively preserved over extended timelines, highlighting the prospective durability of chronically implanted subdural cortical interfaces in the presence of ongoing neurodegenerative illness.
Electrocorticographic, or ECoG, speech brain-computer interfaces show considerable promise for restoring lost communication abilities in patients diagnosed with amyotrophic lateral sclerosis, commonly known as ALS. Despite these technological prospects, a central unresolved question in neural engineering has been whether chronic alterations in signal quality over extended implantation intervals are accompanied by a corresponding loss of task-relevant neural information. To investigate this issue directly, the authors tracked signal characteristics alongside speech decoding metrics over 25 months in a clinical participant with ALS to establish how high-gamma band activity evolves over time and whether any observed neural changes impair offline speech decoding systems.
The clinical research study, registered under the ClinicalTrials.gov identifier NCT03567213, evaluated an individual presenting with slowly progressive bulbar variant ALS. The neurosurgical implementation involved placing two 8 × 8 subdural ECoG grids directly over the participant’s left sensorimotor cortex. Over the full 25-month observational period, the study participant regularly performed an overt syllable-repetition task comprising 12 distinct consonant-vowel tokens. These experimental sessions were conducted with synchronized electrocorticographic monitoring and acoustic audio recording to track neural and vocal outputs concurrently.
To characterize the neurophysiological recordings and vocal changes across time, the investigators quantified three distinct neural metrics focusing on high-gamma activity spanning the 70 to 170 Hz range: the high-gamma activation ratio, the spectral signal-to-noise ratio, and peak z-scored high-gamma responses. Concurrently, vocal motor acoustics were monitored using the triangular vowel space area metric. Longitudinal data trajectories and the timing of physiological transitions were evaluated using generalized additive models and piecewise statistical models, providing a framework to trace how neural representations and acoustic output shifted across hundreds of days.
The statistical models identified a shared, common transition point occurring approximately 180 days following the subdural grid implantation. During the initial period before this 180-day milestone, all three quantified neural signal metrics exhibited upward trajectories. After this transition occurred, the individual neural metrics diverged in their longitudinal paths. The high-gamma activation ratio remained statistically stable at −0.003 dB per 100 days (P = 0.736), whereas active signal-to-noise ratio underwent a pronounced decline of −0.473 dB per 100 days (P < 10^−60). Simultaneously, peak z-scored high-gamma responses continued to increase at a slow rate of 0.067 per 100 days (P < 10^−20), while speech acoustics evaluated via triangular vowel space area demonstrated mild nonlinear reductions.
To evaluate functional communication potential, offline speech decoding was implemented using a modified EEGNet-SE decoder architecture evaluated across 11 chronological rolling train-validation-test splits. Across a total of 726 held-out train-test evaluations, the decoding framework achieved an average 12-syllable classification accuracy of 74.6 percent, far exceeding the theoretical chance performance level of 8.33 percent. Crucially, decoders trained on post-transition data showed no statistically significant decrease in accuracy as the time interval between training and testing data lengthened, exhibiting a minimal trajectory of −0.28 percent per 100 days (P = 0.454) that confirmed strong cross-temporal generalization across the chronic monitoring window.
Detailed analyses of electrode saliency, cross-temporal generalization, and neural representational geometry revealed that higher neural class separability consistently predicted higher held-out decoding accuracy. Furthermore, shuffle-corrected decoded mutual information remained reliably above chance levels without demonstrating any post-transition temporal decline. The authors concluded that the widespread decline observed in active spectral signal-to-noise ratio did not reflect a deterioration of speech-related population neural information, demonstrating that critical spatial decoding patterns, neural class structures, and offline syllable representations can persist durably in a chronic ECoG speech brain-computer interface.
Subject of Research: Longitudinal stability of electrocorticographic speech decoding in amyotrophic lateral sclerosis
Article Title: Stable speech BCI performance during slow progression of ALS: a longitudinal ECoG study
Article References: Ouyang, Z., Walmsley, K., Luo, S., Tippett, D., Wyse-Sookoo, K., Fifer, M., Vansteensel, M. J., Angrick, M., Ramsey, N., & Crone, N. E. (2026). Stable speech BCI performance during slow progression of ALS: a longitudinal ECoG study. Journal of NeuroEngineering and Rehabilitation. https://doi.org/10.1186/s12984-026-02119-y
Image Credits: AI Generated
DOI: 10.1186/s12984-026-02119-y
Keywords: amyotrophic lateral sclerosis, electrocorticography, speech brain-computer interface, high-gamma activity, offline speech decoding, sensorimotor cortex, subdural ECoG grids, syllable repetition task, signal-to-noise ratio, activation ratio, triangular vowel space area, EEGNet-SE decoder, cross-temporal generalization, neural class separability, decoded mutual information, bulbar variant ALS
News Source: Cassandra Pierce. (November 17, 2026). Speech BCI Retains Stable Syllable Decoding Over 25 Months in ALS. Scienmag.



