Bayesian time warping for data-efficient communication brain-computer interfaces

About This Grant

PROJECT SUMMARY Paralysis due to spinal cord injury, stroke, or amyotrophic lateral sclerosis (ALS) can lead to debilitating communication deficits. Implanted brain-computer interfaces (BCIs) are a promising approach to treat these patients. BCIs leverage neural activity to create a desired computer output, such as text. To characterize the relationship between the neural data and the computer output, a decoder is trained on data from many repeated trials. Unfortunately, this training trial burden limits the practical utility of communication BCIs in many patients. There are several contributing factors to this training burden. First, there is an incomplete understanding of the neural codes which underlie complex, skilled behaviors in humans such as handwriting or speech. Second, standard decoders rely on neural network architectures which are extremely flexible but require a substantial amount of training data to achieve acceptable predictive accuracy. Communication BCIs are often based solely on intentions of motion, which creates an additional technical challenge. Due to unobservable variability in the timing of patients’ intentions from trial-to-trial, data-driven methods for aligning neural activity across trials can substantially aid in the analysis of these datasets. I have developed Bayesian time warping for this purpose, a neural activity alignment approach which learns a probability distribution over possible alignments for each trial and response profiles for each neuron based on the observed data. In this project, I propose that the uncertainty estimates generated by this method will provide insights into approaches that can improve the data-efficiency of communication BCIs. To determine if these insights can be generalized across distinct BCI strategies, I will analyze data from two different communication BCI approaches: one which decodes characters from attempted handwriting, and another which decodes phonemes from attempted speech. In Aim 1, I will use probabilistic clustering methods on the model outputs to determine if there are subpopulations of electrodes aligned to distinct features of communication, such as attempted movement onset versus character/phoneme-specific movements, and test if subpopulation-informed BCI decoders require less training data to achieve acceptable performance. In Aim 2, I will use the model outputs to inform the creation of synthetic training datasets of varying size with varying degrees of outlier corruption and systematically characterize the robustness and data-efficiency of decoding architectures ranging in complexity. This project will discover avenues to substantially ease the training trial burden, and therefore, advance the practicality of BCI usage for a larger number of paralyzed patients.