Degradation-aware neural imputation: Advancing decoding stability in brain machine interfaces

Abstract

Neural signal degradation poses a significant challenge in maintaining stable performance when decoding motor tasks using multiunit activity (MUA) and local field potential (LFP) signals in the implantable brain machine interface (iBMI) applications. Effective methods for imputing degraded or missing signals are essential to restore neural signal integrity, thereby improving decoding accuracy and system robustness over long-term recordings with fluctuating signal quality. This study introduces a confidence-weighted Bayesian linear regression (CW-BLR) approach to impute neural signals affected by degradation, enhancing the robustness and consistency of decoding. The performance of CW-BLR was compared to traditional methods—mean imputation (Mean-imp) and Gaussian-mixture-model-based expectation–maximization (GMM-EM)—using a kernel-sliced inverse regression (kSIR) decoder to evaluate decoding outcomes. Four Wistar rats were trained to perform a forelimb-reaching task while neural activity (MUA and LFPs) was recorded over 27 days. CW-BLR imputed signals degraded during days 8–27. Decoding performance was evaluated using kSIR and compared with Mean-imp and GMM-EM. CW-BLR demonstrated superior performance by effectively preserving both temporal and spatial dependencies within the neural signals. CW-BLR-imputed data significantly improved decoding accuracy over traditional imputation methods, with the kSIR decoder showing consistently higher performance, particularly in maintaining signal quality from the degraded period. CW-BLR offers a robust and effective imputation framework for iBMI applications, addressing signal degradation challenges and maintaining accurate decoding over prolonged recordings. By utilizing confidence-based quality metrics, CW-BLR surpasses traditional methods, providing stable neural decoding across fluctuating signal quality scenarios.