Stable speech BCI performance during slow progression of ALS: A longitudinal ECoG study

Background Electrocorticographic (ECoG) speech brain-computer interfaces (BCIs) show promise for restoring communication in amyotrophic lateral sclerosis (ALS), but the long-term stability of speech-related neural signals and decoding performance during disease progression remains unclear. We tracked signal characteristics and decoding over 25 months in a participant with ALS to determine how high-gamma (HG, 70–170 Hz) activity changes over time and whether these changes affect offline speech decoding. Methods We implanted two 8×8 subdural ECoG grids over left sensorimotor cortex (SMC) in a participant with slowly progressive bulbar variant ALS. Across 25 months, the participant performed an overt syllable-repetition task (12 consonant-vowel tokens) during simultaneous ECoG and audio recording. We quantified HG activation ratio (ActR), spectral signal-to-noise ratio (SNR; HG/HF, where HF = 300–499 Hz), and peak z-scored HG responses. Speech acoustics were evaluated using first/second formants (F1/F2) and the triangular vowel space area (tVSA). Offline EEGNet-based decoders were assessed in two stages: models trained on post-implant months 1–6 were tested on months 7–25, while models trained on stabilized data (months 7–11) were tested on the remaining period (months 12–25). Electrode-level saliency assessed spatial contributions to decoding. Results Acoustic analyses showed a significant reduction in tVSA over two years (-44.6 Hz²/day; P  < 10 ⁻ ⁷), consistent with mild intelligibility decline. Neural metrics (ActR and SNR) followed a biphasic trajectory: increasing during the first 6 months, after which ActR stabilized (0.041%/day; P  = 0.13), and SNR declined gradually (-0.46%/day, P  < 10 ^− 4 ). The model trained on months 1–6 achieved 55.7% accuracy (chance: 8.33%), but performance declined over time (-0.019%/day; P  = 2.1×10 ⁻ ⁴). Conversely, the model trained on months 7–11 achieved higher accuracy (65.9%) on subsequent data with no significant temporal decline ( P  = 0.23). Conclusions Speech-related HG features exhibited an initial unstable period followed by a long-term gradual SNR reduction, potentially reflecting disease progression. Models trained after signal stabilization generalized robustly to data recorded over a year later. These findings confirm that despite reduced absolute HG power and mild acoustic degradation of speech, cortical features remain stable enough to support durable ECoG speech BCIs without frequent recalibration. These findings will motivate future adaptive calibration algorithms that account for slow signal changes while leveraging stable spatial representations in ventral SMC. ClinicalTrials.gov Identifier NCT03567213