The Influence of Stimulus Properties, Complexity, and Contingency on the Stability and Variability of Ongoing and Evoked Activity in Human Auditory Cortex A. A. Ioannides,* , † J. G. Taylor,* , ‡ L. C. Liu,* J. Gross,* and H. W. Mu ¨ ller-Ga ¨ rtner* *Institute of Medicine, Research Center Ju ¨ lich, D-52425 Ju ¨ lich, Germany; †Department of Physics, The Open University, Milton Keynes MK7 6AA, United Kingdom; and ‡Department of Mathematics, King’s College, Strand, London WC2R 2LS, United Kingdom Received September 12, 1997 The real-time, single-trial activity in the human auditory cortex was extracted from magnetoencepha- lographic signals. A predictor of single-trial activity was defined as the sum of the average response and a mean-free base level computed over a range of base times. For simple stimuli the residual (predicted– actual) activity had a stimulus-independent oscilla- tory (10 Hz) component. This component was larger and more durable in trained subjects, reaching satura- tion only in the most trained of the five subjects studied (S1). Changes in variability and associated reduction of the absolute value and duration of the oscillations were evident in experiments with stimuli loaded with information, saliency, or task contingency. Repetition reintroduces stimulus-independent oscilla- tions very slowly. For S1, after training, the stimulus- independent oscillations were reestablished in the auditory cortices to the level seen for simple stimuli, except for the time periods and in the hemisphere associated with the combination of task demands and stimulus processing.  1998 Academic Press INTRODUCTION Much of the progress in electrophysiology, at the level of single cell for local field potentials and for macromea- sures mapped with electroencephalography (EEG) and magnetoencephalography (MEG) (Ha ¨ma ¨la ¨ inen et al., 1993) has been built on methods relying on averaging. Recent developments, however, make it possible to study single-trial variability at the level of the single unit (Gur et al., 1997) or by using optical imaging (Grinvald et al., 1982) from a small area (2  2 mm) of cortex (Arieli et al., 1996). At the even larger scale sampled by noninvasive electrophysiology, modern MEG systems provide a very powerful tool for single-trial analysis (Ioannides, 1994). Averaging remains, how- ever, widely used, supported by the folklore that averag- ing removes the ever-present ‘‘physiological noise’’ from the part of the response time-locked to the stimulus. The relevance of this assumption nowadays extends beyond electrophysiology: increasingly more studies attempt to combine results obtained from the analysis of averaged electrophysiological data with results from functional brain imaging like PET and fMRI (Beis- teiner et al., 1995; Snyder et al., 1995; Simpson et al., 1995). The need to couple EEG and MEG remains even as fMRI temporal resolution improves (Busatto et al., 1997; Buckner et al., 1996), as long as one wishes to study directly the millisecond-by-millisecond sequence of events which are inaccessible through hemodynamic- based measures because of the slow nature of these processes. Neither fMRI nor PET relies on averaging of quantities which can have positive and negative compo- nents, and hence there is no cancelation because of latency jitter in the response. A prerequisite for combin- ing PET and fMRI with average MEG or EEG is that the sequence of peaks and troughs in the average waveform is an accurate record of the activity se- quences in single trials. Single-trial MEG analysis has repeatedly challenged this assumption, suggesting in- stead that the influence of an external stimulus leads to a reorganization of ongoing activity with large latency jitter. Our earlier studies (Liu and Ioannides 1996; Liu et al., 1998) suggested that the single-trial activity in- duced in the auditory cortex by a simple tone cannot be treated as a deterministic response emerging from a random noisy background. Recent optical imaging stud- ies (Arieli et al., 1996) seemed to reach the opposite conclusion: the real-time, single-trial, local evoked re- sponse can be adequately described by a linear superpo- sition of a stimulus-driven deterministic response and the preceding ongoing activity. We have therefore looked again at the question of what constitutes ‘‘signal’’ and what constitutes ‘‘noise’’ along lines similar to those of Arieli et al. The main difference in methods is that we NEUROIMAGE 8, 149–162 (1998) ARTICLE NO. NI980353 149 1053-8119/98 $25.00 Copyright  1998 by Academic Press All rights of reproduction in any form reserved.