# High-gamma and beta bursts in the left supramarginal gyrus can differentiate verbal memory states and performance

## TL;DR

This study shows that high-gamma and beta brain signals in the left supramarginal gyrus can help distinguish between memory states and performance during verbal recall.

## Contribution

The novel contribution is using high-gamma and beta bursts in the LSMG to classify memory states and performance with machine learning.

## Key findings

- CNNs using high-gamma and beta burst data from the LSMG could classify memory states with AUROC scores above 0.52 in 62 experiments.
- A generalized linear model showed a significant interaction between signal-to-noise ratios of HG and beta bursts and recall probability.
- CNNs could distinguish encoding from scrambled recall epochs, suggesting these signals encode memory information.

## Abstract

The left supramarginal gyrus (LSMG) contributes to attentional allocation for memory encoding and may also reflect memory state and performance. Given the roles of high-gamma and beta bursts in cognition and memory, this proof-of-concept study investigated whether these signals within the LSMG could classify memory state and performance.

Using secondary data from 103 epilepsy patients undergoing presurgical iEEG evaluation, we analyzed 141 delayed verbal free recall experiments. Intracranial EEG (iEEG) data, recorded solely from LSMG electrode contacts, were processed to create two-dimensional (2D) tensors of convolved high-gamma (HG), and beta (15–40 Hz) burst activity. Convolutional neural networks (CNNs) were trained and cross-validated on these 2D tensors to classify memory state (encoding versus recall) and performance (remembered versus forgotten items) within subjects.

The latter CNN, used to label subsequently recalled words based on iEEG recorded during the encoding epoch, performed at or below chance in 79 of the 141 experiments. In all but 3 of these 79 experiments, the iEEG was contaminated or low amplitude. In the other 62 experiments this CNN labeled recalled words with an area under the receiver operating curve (AUROC) score of greater than 0.52. A generalized linear model explained the variance of the AUROC score for labelling recalled words correctly in these 62 experiments (n = 62, d.f. = 20, F = 1.7, p = 1 × 10−4). The most significant term in the model was a positive interaction between (1) mean HG burst signal to noise ratio; (2) mean beta burst signal to noise ratio; (3) the number of electrode contacts in the LSMG; and (4) recall probability (t = 3.04, p = 0.006). We identified 14 experiments that labeled subsequently recalled words during encoding with an AUROC score greater than 0.6. To address over-training, we also trained and then tested the CNN on distinct datasets in four subjects. In most of these experiments CNN performed better than chance. We also found that a CNN utilizing 2D tensors of HG and beta bursts could distinguish encoding from scrambled recall epochs.

This work indicates LSMG is a memory hotspot and that HG and beta bursts may serve as temporal memory information packets or signify attention related to memory.