Neural voice activity detection with high-gamma ECoG signal correlation structure using a chronically implanted brain-computer interface in an individual with ALS

Abstract

Chronically implanted brain-computer interfaces (BCI) for speech have demonstrated restoration of speech communication for people with severe motor impairment. Yet maintaining stable long-term performance remains a translational challenge. We investigated whether correlation and partial correlation features of electrocorticographic (ECoG) high-gamma activity (HG-C, HG-PC) could improve robustness compared with high-gamma log-power (HGLP) features for neural voice activity detection (NVAD). Using open-source BCI data from an individual with amyotrophic lateral sclerosis performing a syllable-repetition task, long short-term memory (LSTM) models were trained separately on HGLP, HG-C, and HG-PC features, and evaluated across sessions spanning six months. HG-C and HG-PC achieved comparable or superior NVAD performance to HGLP with smaller long-term averaged loss (HGLP: -17%; HG-C: -9%; HG-PC: -12%) and smaller long-term worst case loss (HGLP: -24%; HG-C: -8%; HG-PC: -16%). Under a simulated local contiguous pattern of neural signal loss, both HG-C and HG-PC outperformed HGLP on averaged long-term loss (HGLP: -22%; HG-C and HG-PC, -10%) and worst-case long-term loss (HGLP: -30%; HG-C: -11%; HG-PC: -13%). The results show that high-gamma correlation-based features captured comparatively more spatially distributed and stable neural speech representations. With further refinement and validation, correlation-based feature representations may contribute to robust longitudinal speech decoding with implanted BCIs.