FIGURE 1.

The proposed decoding framework for handling channel disruptions. (A) Raw voltage data and impedance measurements recorded from the implanted electrode array will be delivered to the statistical process control algorithm, where channels with disruptions will be identified. The discovery of disruptions will cause the input from the affected channels to be masked, or zeroed, before being sent to the decoder. This will also trigger an unsupervised update to readjust the weights to the missing input. (B) The statistical process control algorithm to detect disrupted channels is described. The four steps in this process are (1) transforming raw data into four array-level metrics useful for signal monitoring, (2) creating control charts for each of the metrics, (3) using the control charts to flag sessions with potential disruptions, and (4) performing Grubb’s test to determine outlying channels on the flagged sessions. (C) The decoder framework, including a masking layer, an LSTM with 80 hidden units, 25 convolutional filters, a fully connected layer, and finally a 5-unit layer corresponding to the output.