Yoon, S., Avansino, D. T., Madugula, S., Levin, A. D., Fan, C., Abramovich Krasa, B., Singh, A., Vo, C. et al.
This paper validates the real-time applicability of deep ensemble methods for improving speech decoding in brain-computer interfaces (BCIs) and proposes a computationally efficient alternative.
Speech decoding in BCIs suffers from high error rates, limiting practical communication. While deep ensemble methods significantly improve accuracy, they lack real-time testing, require substantial computational resources, and their performance under clinical constraints is poorly understood.
The study conducted the first closed-loop (real-time) test of deep ensembles on a participant with bilateral intracortical microelectrode arrays. It systematically analyzed performance dependencies on baseline error rate, training dataset size, and ensemble size. A computationally efficient pseudoensembling approach based on test-time augmentation, requiring only a single base decoder, was introduced.
The deep ensemble approach reduced word error rate from 33.7% to 26.0% on a large-vocabulary task in real-time. The study provided a comprehensive assessment of performance under various clinically relevant constraints and resource-accuracy tradeoffs. The proposed pseudoensembling method maintains accuracy gains while drastically reducing computational burden, offering a practical pathway for broader clinical adoption of speech BCIs.