Figure 10.

Recording disruptions impair motor intention decoding in a BMI clinical trial. The performance of three deep neural network decoder variants were evaluated for a four-movement motor imagery task over the span of 1 year. Accuracies are plotted as a function of the number of days since the end of the neural network training period. Lines denote a LOESS smoothing curve to visualize the data trends. The fixed neural network (fNN, cyan circles) decoder parameters remained unchanged for the duration of the evaluation. The other networks were updated each session with data from a preceding recording block with either explicit training labels (sNN, green diamonds) or with labels predicted by the decoder (uNN, orange squares). Both the sNN and uNN can adapt to daily changes in recording conditions and thus outperform the fNN. On days 238 and 266 (darkened data points) a disruption caused a substantial drop in accuracy for all decoding models. These dates corresponded to times when the participant had undiagnosed infections. Though the adaptive decoders are better able to compensate for this disruption, additional algorithmic interventions are needed to prevent sharp declines in decoding accuracy. Figure adapted with permission from Schwemmer et al. (2018). Copyright 2018, Springer Nature.