Interpreting neural decoding models using grouped model reliance

Abstract The application of machine learning algorithms for decoding psychological constructs based on neural data is becoming increasingly popular. However, there is a need for methods that allow to interpret trained decoding models, as a step towards bridging the gap between theory-driven cognitive neuroscience and data-driven decoding approaches. The present study demonstrates grouped model reliance as a model-agnostic permutation-based approach to this problem. Grouped model reliance indicates the extent to which a trained model relies on conceptually related groups of variables, such as frequency bands or regions of interest in electroencephalographic (EEG) data. As a case study to demonstrate the method, random forest and support vector machine models were trained on within-participant single-trial EEG data from a Sternberg working memory task. Participants were asked to memorize a sequence of digits (0–9), varying randomly in length between one, four and seven digits, where EEG recordings for working memory load estimation were taken from a 3-second retention interval. Present results confirm previous findings in so far, as both random forest and support vector machine models relied on alpha-band activity in most subjects. However, as revealed by further analyses, patterns in frequency and particularly topography varied considerably between individuals, pointing to more pronounced inter-individual differences than reported previously. Author summary Modern machine learning algorithms currently receive considerable attention for their predictive power in neural decoding applications. However, there is a need for methods that make such predictive models interpretable. In the present work, we address the problem of assessing which aspects of the input data a trained model relies upon to make predictions. We demonstrate the use of grouped model-reliance as a generally applicable method for interpreting neural decoding models. Illustrating the method on a case study, we employed an experimental design in which a comparably small number of participants (10) completed a large number of trials (972) over multiple electroencephalography (EEG) recording sessions from a Sternberg working memory task. Trained decoding models consistently relied on alpha frequency activity, which is in line with existing research on the relationship between neural oscillations and working memory. However, our analyses also indicate large inter-individual variability with respect to the relation between activity patterns and working memory load in frequency and topography. Taken together, we argue that grouped model reliance provides a useful tool to better understand the workings of (sometimes otherwise black-box) decoding models.