To isolate brain activity that may reflect effective cognitive processes during the study phase of a memory task, cognitive neuroscientists commonly contrast brain activity during study of later-remembered versus later-forgotten items. This "subsequent memory effect" method has been described as identifying brain activity "predictive" of memory outcome. However, the modern field of machine learning distinguishes between descriptive analysis, subject to overfitting, and true prediction, that can classify untrained data. First, we tested whether classic event-related potential signals were, in fact, predictive of later old/new recognition memory (N = 62, 225 items/participant); this produced significant but small predictive success. Next, pattern classification of the multivariate spatiotemporal features of the single-trial EEG waveform also succeeded in predicting memory. However, the prediction was still small in magnitude. In addition, topographic maps suggested individual differences in sources of predictive activity. These findings suggest that, on average, brain activity, measured by EEG, during the study phase is only marginally predictive of subsequent memory. It is possible that this predictive approach will succeed better when other experimental factors known to influence memory outcome are also integrated into the models.NEW & NOTEWORTHY For both basic and applied reasons, an important goal is to identify brain activity present while people study materials that enable us to predict whether they will remember those materials. We show that this is possible with the conventional event-related potential "subsequent-memory-effect" signals as well as with machine learning classifiers, but only to a small degree. This is in line with behavioral research, which supports many determinants of memory apart from the cognitive processes during study.
Keywords: event-related potential; pattern classification; prediction; recognition memory; subsequent memory effect.