Evaluating EEG activity as predictive of memory through classifiers

Brain-activity measures have the potential to provide powerful new constraints on memory models. With classifier-based approaches, one can identify signals, derived from a training-set data, that can predict memory outcome on test-set data. Advancing beyond descriptive methods, the classifier-based approach could identify brain-activity features that are more likely to be behaviourally relevant, rather than spectator or performance-irrelevant processes. Instead of chasing down optimal classification, we take a systematic approach to evaluate this, and to identify where improvements to classifier approaches could be made. We start with univariate event-related potential measures that have previously been implicated in recognition-memory study and matching processes (study: LPC and slow wave; test: FN400 and LPP). In 64 participants performing old/new verbal recognition, univariate measures predicted memory-accuracy with small, but significant, success (95% CI AUC = study: [0.51 0.54]; test: [0.52 0.55]; chance=0.5). Multivariate, LDA and SVM spatio-temporal classifiers performed better (study: [0.52 0.56]; test: [0.55 0.60]), suggesting the importance of other features beyond these previously identified ERP features. Overall success rates remained remarkably low, but this is in line with results from other related published approaches. However, AUC approached 0.7 for high-performing participants. Addressing individual differences in motivation/engagement, or titrating difficulty, may lead to high classification success. Future approaches should also incorporate the myriad known behavioural factors that determine memory outcome but are absent in brain activity during study or test of a particular item.