Abstracts

Rationale: A significant part of EEG interpretation consists in recognizing motifs that consistently recur. At a basic level, these motifs consist of waveforms within single EEG recording channels. The collection of all such waveforms can be regarded as a dictionary in which the words are recurrent single-channel EEG waveforms. While several dozen words in this dictionary have been identified manually over the years, how many distinct words may be said to exist within the EEG, within a given subject or across a population is unknown. Automated computational methods may in principle greatly accelerate construction of the EEG dictionary, and may identify previously unknown words. In this work we developed novel techniques for efficiently searching massive time-series data sets to estimate the size and statistical structure of a putative intracranial EEG dictionary. Methods: We analyzed data from 471 single-channel, 30 minute-long segments of intracranial EEG recordings from 4 different patients monitored for epilepsy surgery, to compute words varying in length from 40-3000 milliseconds. Recurrent waveforms were identified by comparing each possible waveform of a given length to all others of the same length, and retaining those which surpassed a psychophysically determined similarity threshold. While a brute-force search requires approximately one trillion waveform comparisons, the methods presented herein permitted an exact solution while allowing the vast majority of these comparisons to be bypassed, rendering the problem computationally tractable. Results: The EEG dictionary computed for each individual patient was largely complete, i.e. merging the dictionaries from other patient s EEGs added very few additional words. The total number of words within the final, complete dictionary contained approximately 114,233 words, a relatively small fraction of all possible words. As for many actual spoken languages (e.g. English), the number of occurrences of each word within the dictionary roughly obeys a Zipf distribution, with the most frequent 5% of words accounting for approximately 80% of all words within the EEG, i.e. roughly 9 times more frequently than the remaining 95% of all words combined. A control dictionary, computed from artificial data designed to match the power spectra of each patient s EEG but to contain no higher order statistical structure, was 1.5 times as large. 54% of words were common between the two dictionaries. Conclusions: The intracranial EEG appears to be composed of a relatively restricted collection of waveforms, the vast majority of which are universal across patients. While roughly half of the words in the EEG dictionary follow directly from the structure of the EEG power spectrum, the remaining words reflect more complex dynamical processes underlying cerebral activity. This novel approach to decomposing the EEG into a set of recurrent waveforms, and the computational methods developed for this purpose, hold promise for many future data mining and pattern recognition applications in the computational analysis of EEG.