Introduction: Closed-loop brain stimulation devices have demonstrated promise in improving episodic memory. However, such devices rely on accurate classification of real-time electrophysiological correlates of memory-related brain activity as either conducive or non-conducive to successful episodic encoding. We compared the classification accuracy of a recurrent neural networks (RNNs) paradigm to that of our group's previously reported highest performing machine learning classifier-support-vector machines (SVM) with t-distributed stochastic neighbor embedding (tSNE). Also, by withholding information recorded from various brain regions from the SVM/tSNE classifier, we predicted the impact of tissue resection during epilepsy surgery on episodic memory performance following the operation.
Methods: Fifteen patients with medically refractory epilepsy were implanted with intracranial electrodes. All had contacts in the dominant hemisphere of five common brain regions-the hippocampus, precuneus, posterior cingular gyrus, lateral temporal cortex, and posterior lateral temporal cortex. While implanted, patients participated in an episodic memory task (free recall). EEG signals from memory encoding periods were organized by region (the five aforementioned), frequency band (six logarithmically spaced bands ranging from 2.5-100 Hz), and time windows (six sequential windows within the 1800 ms encoding events) and fed to both classifier types.
Results: RNNs significantly outperformed the SVM/tSNE model as demonstrated by their respective area under the receiver operating characteristic curve values (RNNs = 0.72, SVM/tSNE = 0.68, p=0.0026). When oscillatory information from each of the five regions was withheld from the SVM/tSNE classifier in turn, hippocampal information loss resulted in the greatest decline in classification accuracy (p=0.0039).
Conclusions: Improving the classification accuracy of closed-loop stimulation devices through use of RNNs has the potential to boost the memory improvements seen with their use. Also, harnessing machine learning to quantify the relative importance of each region to classifier performance provides a novel way of predicting memory loss after tissue resection in epilepsy surgery.
Patient Care: Closed-loop stimulation devices for memory improvement have shown promise, but their success remains predicted on the success of their chosen classification technique. Maximizing classification accuracy will be an integral part to their potential future success in helping patients with neurocognitive deficits stemming from a range of pathologies recover some memory functionality. These devices may become more commonplace soon, and it will be important to maximize their performance before their potential larger-scale implementation. Also, using classifiers to predict memory loss after epilepsy surgery is an exciting technique, but one still in need of prospective validation. Sharing this with the CNS allows more institutions to potentially join in this process. If validated, this is a simple, understandable metric that can be brought to the bedside to aid patients in their health-care decisions regarding epilepsy surgery. Empowering patients with information on their condition and prognosis is valued.
Learning Objectives: By the conclusion of this session, participants should be able to:
1. note the improvement in classification accuracy deep learning affords closed-loop stimulation devices
2. understand the potential use of recurrent factor elimination and machine learning to predict memory loss following epilepsy surgery