INMED/TINS special issue Analysis of dynamic brain oscillations: methodological advances Michel Le Van Quyen 1 and Anatol Bragin 2 1 LENA-CNRS UPR640, Universite ´ Pierre et Marie Curie, Ho ˆ pital de la Salpe ˆ trie ` re, 75651 Paris Cedex 13, France 2 Department of Neurology and Brain Research Institute, David Geffen School of Medicine at UCLA, Los Angeles, CA 90095-1769, USA In recent years, new recording technologies have advanced such that, at high temporal and spatial resol- utions, oscillations of neuronal networks can be identified from simultaneous, multisite recordings. However, bec- ause of the deluge of multichannel data generated by these experiments, achieving the full potential of parallel neuronal recordings also depends on the development of new mathematical methods that can extract meaningful information relating to time, frequency and space. Here, we aim to bridge this gap by focusing on up-to-date recording techniques for measurement of network oscil- lations and new analysis tools for their quantitative asse- ssment. In particular, we emphasize how these methods can be applied, what property might be inferred from neuronal signals and potentially productive future direc- tions. This review is part of the INMED and TINS special issue, Physiogenic and pathogenic oscillations: the be- auty and the beast, derived from presentations at the annual INMED and TINS symposium (http://inmednet. com). Introduction: the complex web of neuronal oscillations Advances in our understanding of neural systems go hand-in-hand with improvements in experimental tech- niques used to study these systems. From the rapid growth in biotechnology, multisite recording techniques now enable monitoring of ensemble oscillations in great detail by sim- ultaneously recording local neuronal activity from a large number of network locations [1] (Box 1). What emerges from these parallel neuronal recordings is a rich picture of brain dynamics, in which locally generated oscillations occur transiently in different brain regions and mediate coordi- nated interactions within and between different neuronal subsystems [2–7]. A useful analogy is the Internet, in which geographically distant computers briefly transfer data to each other within transient assemblies that are formed on a static network of hard-wired connections [6]. Following the Internet analogy, neuronal oscillations define short temporal windows for flexible communication between widely distributed neuronal ensembles [8,9]. Multichannel measurements that sample cortical space and time are, therefore, necessary to gain a detailed understanding of these complex spatiotemporal behaviors. However, as a result of improvements in the capabilities of the measuring apparatus, in addition to the growth of storage capacity, the data sets generated by these exper- iments are increasingly large and more complex. There- fore, analysis of such multichannel data is an important piece of the associated research program on oscillations. Standard computer software [e.g. MATLAB (Mathworks, http://www.mathworks.com); Box 2] now makes math- ematical analysis relatively easy, and modern computers have the necessary speed and memory capacity to apply these techniques to large data sets (Box 3). Nevertheless, although traditional methods of data analysis are useful for many purposes, the growing complexity of neuroscien- tific experiments, often examining subtle frequency changes on a comparatively fine timescale, requires careful attention to more sophisticated methods of data analysis. Here, we discuss fundamental issues of data analysis that face researchers characterizing the dynamic oscil- lations observed in continuous signals, such as local field potentials. We illustrate general points using the problems of describing the evolution of a neuronal oscillation in time–frequency and assessing time-varying synchroniza- tions between several oscillations in space. We also refer to Internet sites from which it is possible to obtain software codes for putting into practice the different methods reviewed herein. Time–frequency structures of dynamic oscillations The key to extracting information from a set of measurements is to display those measurements in another equivalent representation in which their information con- tent becomes obvious. Often, the key to extracting infor- mation is to switch from a temporal domain to a frequency domain. The first such transformation in wide use is the Fourier transform, providing spectral power that identifies the amplitudes of sine functions of various frequencies that exist throughout the entire duration of the signal (Figure 1a). In practice, neuronal oscillations are dynamic, with frequency content that changes over time, and the Fourier transform elucidates the spectral content of the signal; however, it provides no information regarding the point in time at which those spectral components app- ear. In other words, not only must the frequency components Review TRENDS in Neurosciences Vol.30 No.7 Corresponding author: Le Van Quyen, M. (lenalm@ext.jussieu.fr). Available online 7 June 2007. www.sciencedirect.com 0166-2236/$ – see front matter ß 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.tins.2007.05.006