A spatio-spectral mulitmodal approach for movement decoding in patients with Parkinson's disease

VICTORIA PETERSON; TIMON MERK; BUSH, ALAN; VADIM NIKULIN; ANDREA A. KÜHN; WOLF-JULIAN NEUMANN; MARK RICHARSON

Conferencia; 7th Annual BRAIN initiative meeting; 2021

The application of machine learning to intracranial signal analysis has the potential to revolutionize deep brain stimulation (DBS) by personalizing therapy to dynamic brain states, specific to symptoms and behavior. Machine learning methods can allow behavioral states to be decoded accurately from intracranial local field potentials to trigger an adaptive DBS (aDBS) system, closing the loop between patients’ needs and stimulation pattern. Most decoding pipelines for aDBS are based on single channel frequency domain features, neglecting spatial information available in multichannel recordings. Such features are extracted either from DBS lead recordings in the subcortical target or from electrocorticography (ECoG). To optimize the simultaneous use of both types of signals, we developed a supervised online-compatible decoding pipeline based on multichannel and multimodal recordings, using data obtained from 11 patients with Parkinson's disease performing a hand movement task during DBS surgery, in which a research ECoG electrode was temporary placed. Spectral and spatial features were extracted using filter-bank analysis and spatial pattern decomposition. The learned spatio-spectral features were used to train a generalized linear model with sparse regularized regression. We found that, dimensionality reduction improved decoding for multimodal recording compared to either unimodal approach, for both contralateral and ipsilateral movements (decoding improvement up to 15%). The prediction value was inversely correlated with both the UPDRS score and the distance of the ECoG electrode position to the hand knob motor cortex. The analysis of the selected features revealed that ECoG signals contribute more significantly to the decoding performance than subcortex signals. We also found that decoding was successful using 100 ms time windows, a time delay that is well-suited for aDBS applications. This novel application of spatial filters to decode movement from combined cortical and subcortical recordings is an important step toward the use of machine learning for the construction of intelligent aDBS