Full-length reviewEEG alpha and theta oscillations reflect cognitive and memory performance: a review and analysis
Introduction
In a physiological sense, EEG power reflects the number of neurons that discharge synchronously. Because brain volume and the thickness of the cortical layer is positively correlated with intelligence (e.g., Refs. 12, 165) it is tempting to assume that EEG power too, is a measure that reflects the capacity or performance of cortical information processing. Although it will be argued that this is in principle the case, it must be emphasized that power measurements are strongly affected by a variety of unspecific factors such as the thickness of the skull or the volume of cerebrospinal fluid, by methodological and technical factors (such as interelectrode distance or type of montage) but also by more specific factors such as age, arousal and the type of cognitive demands during actual task performance.
It is the purpose of the present article to show that EEG power is indeed related to cognitive and memory performance, but in a complex and partly non-linear way. Within the alpha frequency range EEG power is positively related to cognitive performance and brain maturity, whereas the opposite holds true for the theta frequency range. Alpha and theta reactivity as well as event-related changes in alpha and theta band power show yet another pattern of results. During actual task demands the extent of alpha power suppression is positively correlated with cognitive performance (and memory performance in particular) whereas again the opposite holds true for the theta band. Here, the extent of theta synchronization is related to good performance. The review which focuses on the theta and alpha frequency range considers first tonic changes in power (such as age related differences in the EEG) and then phasic or event-related changes.
Two basic aspects of memory processes will be distinguished [66]. The first refers to processes of the working memory system (WMS), the second to that of the long-term memory system (LTMS). Probably any cognitive process depends on the resources of both systems. As an example, let us consider an every day cognitive process such as recognizing a familiar object. The basic idea here is that after a sensory code is established, semantic information in long-term memory (LTM) is accessed which is used to identify the perceived object. If the matching process yields a positive result, the object is recognized which in turn leads to the creation of a short-term memory (STM) code. In this case, bottom up pathways are activated which are similar or identical to those which would serve to retrieve information from LTM. This classical explanation of encoding still reflects the current view, which was originally stated by Shiffrin and Geisler [137]: “The process of encoding is essentially one of recognition: the appropriate image or feature is contacted in LTM and then placed (i.e., copied) in STM” (p. 55). Complex cognitive processes such as speaking and thinking may also be described in terms of a close interaction between the WMS and LTMS. The basic difference to the foregoing example is that a sensory code is lacking and that a code is generated in STM which in the case of speaking represents a `plan' of what to say. The codes generated in STM trigger search processes in LTM to retrieve the relevant knowledge about the appropriate semantic, syntactic and articulatory information. This latter idea is similar to Baddeley's concept of working memory, which comprises an attentional controller, the central executive and subsidiary slave systems 3, 4.
In the WMS, encoding has two different meanings, one refers to episodic, the other to sensory-semantic information. The encoding of sensory information (as a process of recognition) always aims at the semantic understanding of perceived information which is first processed in the LTMS. Because of this close relation between sensory and semantic encoding we will use the term sensory-semantic code. The creation of a new code comprises episodic information, which according to Tulving (e.g., [158]), is that type of contextual information which keeps an individual autobiographically oriented within space and time. Because time changes the autobiographical context permanently, there is a permanent and vital need to update and store episodic information. Thus, the formation of episodic memory traces is one of the most important tasks of the WMS.
Cognitive performance is closely related and linked to the performance of the WMS and LTMS. As an example, most intelligence tests comprise subtests measuring memory span (an important function of the WMS) and tests such as judging analogies (which reflect an important aspect of semantic LTMS). The increase in cognitive performance from childhood to puberty as well as the decrease in performance during the late lifespan is, most likely, due or at least closely linked to performance changes in the WMS and LTMS.
With respect to the functional anatomy of memory, there is good evidence that brain structures that lie in the medial temporal lobe (comprising the hippocampal formation) and prefrontal cortex support various functions of the WMS (cf. the reviews in Refs. 102, 140, 141and findings about the contribution of the hippocampal region for the generation of event-related potentials in the human scalp EEG during novelty detection 85, 86). Studies focusing on the ontogeny of human memory indicate that the hippocampus matures relatively early in postnatal life, whereas the prefrontal cortex which is important for the development of an increased memory span matures much later (cf. Refs. 109, 57for reviews). Although many aspects of memory develop early in childhood (up to an age of about 2 or 3 years) it is not yet known when memory is fully matured. As will be discussed in Section 5, there is good evidence that a complex structure of feedbackloops (or `reentrant loops'; [37]) connecting the hippocampus with different cortical regions and the prefrontal cortex in particular may provide the anatomical basis for the WMS. It appears likely that these feedbackloops develop and become increasingly differentiated with increasing age over the entire time span of childhood and possibly early adulthood. The increase in EEG frequency during that life period may (besides other factors) reflect this process of brain maturation. The basic assumption is that the better these feedbackloops become integrated and interconnected with other brain areas, the faster the frequency of EEG oscillations will be.
Alpha is (with the exception of irregular activity in the delta range and below) the dominant frequency in the human scalp EEG of adults. It is manifested by a `peak' in spectral analysis (cf. Fig. 1) and reflects rhythmic `alpha waves' which are known since Berger (e.g., Ref. [10]). The fact that alpha clearly is an oscillatory component of the human EEG has led to a recent `renaissance' in the interest of EEG alpha activity 5, 8, 7, 9.
Frequency and power are closely interrelated measures. Usually, alpha frequency is defined in terms of peak or gravity frequency within the traditional alpha frequency range (f1 to f2) of about 7.5–12.5 Hz. Peak frequency is that spectral component within f1 to f2 which shows the largest power estimate (cf. Fig. 1A). Alpha frequency can also be calculated in terms of gravity (or `mean') frequency which is the weighted sum of spectral estimates, divided by alpha power: (∑(a(f)×f))/(∑ a(f)). Power spectral estimates at frequency f are denoted a(f). The index of summation is in the range of f1 to f2. Particularly if there are multiple peaks in the alpha range (for a classification see e.g., Ref. [39]), gravity frequency appears the more adequate estimate of alpha frequency.
As is well known from animal research, unlike alpha in the human scalp EEG, theta is the dominant rhythm in the hippocampus of lower mammals. Its frequency ranges from about 3 to 12 Hz (e.g., Ref. [95]) and, thus, shows a much wider frequency range than in humans where theta lies between about 4 to 7.5 Hz. Its wide frequency range and large power make it easy to observe frequency and power changes in animals. This is in sharp contrast to the human scalp EEG, where—without the help of sophisticated methods—changes in theta frequency are very difficult or almost impossible to detect. The question, thus, is whether there is a physiological criterion that allows us to decide which frequency marks the transition between alpha and theta oscillations.
Alpha and theta respond in different and opposite ways. The crucial finding is that with increasing task demands theta synchronizes, whereas alpha desynchronizes (cf. the bold in relation to dotted line in Fig. 1A). This fact is documented in reviews on event-related desynchronization (ERD) 122, 67, 68, as well as by a variety of studies using other experimental approaches (e.g., Ref. 45, 46, 95, 99, 131, 133, 129). If EEG power in a resting condition is compared with a test condition, alpha power decreases (desynchronizes) and theta power increases (synchronizes). Classical findings demonstrate that the decrease in alpha power (suppression of the alpha rhythm) can be observed primarily when subjects close their eyes. More recent evidence, however, suggests that attentional and semantic memory demands are powerful factors which lead to a selective suppression of alpha in different `subbands' and that the well described effects of visual stimulation (e.g., eyes open vs. closed) represent just a special class of sensory-semantic task demands (see Section 4and Refs. 72, 75, 78, 79, 80, 81). As already emphasized, encoding of sensory information always aims to extract the meaning of the perceived information which is stored in semantic LTM. Thus, there is a close relation between sensory and semantic encoding.
As Fig. 1 illustrates, that frequency in the power spectra which marks the transition from theta synchronization to alpha desynchronization may be considered the individual transition frequency (TF) between the alpha and theta band for each subject. When using this method to estimate TF, we have found that TF shows a large interindividual variability (ranging from about 4 to 7 Hz) which is significantly correlated with alpha peak frequency [76]. Preliminary evidence for a covariation between theta and alpha frequency was already found by Klimesch et al. [75]and is further documented by Doppelmayr et al. [34]. These findings indicate that theta frequency (as measured by TF) varies as a function of alpha frequency and suggest to use alpha frequency as a common reference point for adjusting different frequency bands not only for the alpha, but theta range as well. For estimating theta power, the individual determination of frequency bands may even be more important because otherwise the effects of theta synchronization are masked by alpha desynchronization particularly in the range of TF (cf. Fig. 1B,C).
Because alpha frequency varies to a large extent as a function of age, neurological diseases, memory performance (see Section 2.1below), brain volume 115, 114and task demands [73], the use of fixed frequency bands does not seem justified. As an example, an elderly subject with bad memory performance may show a peak frequency of 7 Hz or lower [16]. When strictly applying the rule that alpha peak frequency is that spectral component within f1=7.5 and f2=12. 5 Hz which shows maximal power, we would arrive at the conclusion that the obtained frequency indicates theta instead of alpha frequency. Fortunately, as discussed in the previous section, there is a physiological criterion which allows us to answer this question. If EEG power around 7 Hz would desynchronize during a test—as compared to a resting condition (cf. Fig. 1A)—we still would accept that a peak frequency of 7 Hz indicates alpha but not theta frequency.
This example documents the necessity to define the alpha band individually for each subject as that range (f1 to f2) `around' the individual dominant EEG frequency (above the lower delta range) that desynchronizes during task demands. In order to avoid confusions with traditional measures, we use the term individual alpha frequency (IAF) to denote the individual dominant (peak or gravity) EEG frequency (in the range of f1 to f2) of a single subject. The crucial point, of course, is the exact location and individual definition of the frequency limits f1 and f2. For f1, TF is a good estimate, but for f2 an obvious physiological criterion is lacking. An indirect way to solve this problem is to define the frequency limits of the lower alpha band by f1 and IAF and to assign the `remaining' part of the alpha frequency range (which equals [alpha frequency window]−[IAF−f1]) to the upper band. Results from our laboratory indicate that the lower alpha band has a width of about 3.5–4 Hz. Accordingly, the upper alpha band (the frequency range above IAF) is a rather narrow band of 1 or 1.5 Hz, if it is assumed that the alpha frequency window has a width of about 5 Hz. Experimental findings (discussed in detail in Section 4below) indicate that the upper alpha band—defined as a band of 2 Hz above IAF—responds selectively to semantic LTM demands and behaves in a completely different and sometimes opposite way as the lower alpha band (see also the review in Ref. [121]). Furthermore and most importantly, it was found that the lower alpha band (a band of 4 Hz below IAF) reflects different types of attentional demands. Thus, it was broken down into two subbands of 2 Hz each which are termed lower-1 and lower-2 alpha (see Fig. 1 and Section 4).
In summarizing, when using IAF as an anchor point, it proved useful to distinguish three alpha bands (with a width of 2 Hz each), two lower alpha bands (below IAF) and one upper alpha band (above IAF). The theta band is defined as the frequency band of 2 Hz which falls below TF. As for the upper frequency limit of the upper alpha band, there are no clear criteria for the lower frequency limit of the theta band. In any case, however, it is important to emphasize that the use of narrow frequency bands reduces the danger that frequency specific effects go undetected or cancel each other. Thus, broad band analyses must be interpreted with great caution. The implication is that an unbiased estimate of alpha and theta power can be obtained only, if the traditional fixed band analyses are abandoned and if narrow frequency bands are adjusted to the individual alpha frequency of each subject. However, the vast majority of studies use broad, instead of narrow alpha bands (cf. Table 1) which in addition are not adjusted individually.
The suggested definition of frequency bands is based on physiological criteria (such as TF and IAF) and on the functional significance of narrow frequency bands (see Section 4). An alternative way to define EEG frequency bands is to analyze the covariance of spectral estimates by multivariate statistical methods such as factor analysis. The crucial question here is whether the activities in different frequency bands vary independently. With respect to the traditional theta and alpha frequency range, results from factor analyses show at least three independent factors, one for theta, one for the lower alpha band and another for the upper alpha band (see e.g., Ref. [106]and the summary in Ref. [96]). However, the frequency limits vary considerably between studies. As an example, Wieneke (reported in Ref. [96]) found a factor covering the frequency range of 6–9 Hz, which was termed `theta' whereas within that same frequency range, Mecklinger et al. [106]extracted one component which they classified as `lower alpha'. Divergent results from factor analyses are due to the type of power measurements (spectral estimates may be expressed in terms of relative or absolute power), type of derivation (e.g., monopolar [referential], bipolar, Laplacian), electrode location, task type (resting condition with eyes open or closed or performance of some task), the selected sample of subjects, and finally the method used to extract and rotate factors. It also should be emphasized that factor analyses are usually performed on the basis of a correlation matrix which was obtained by correlating spectral estimates over a sample of n subjects. Done in this way, the extracted factors represent average frequency ranges of that particular sample. However, factor analysis could also be used for an individual definition of frequency bands, if the EEG of a single subjects is used and if spectral estimates are correlated over a series of n trials (epochs).
Even if frequency bands are defined individually for each subject, it must be emphasized that EEG frequencies vary between recording sites. As an example, it is well known that alpha waves occur primarily during wakefulness over the posterior regions of the head and can be best seen with eyes closed and under conditions of physical relaxation and mental inactivity [110]. The frequency of alpha waves is faster at posterior and slower at anterior recording sites. It would, thus, be desirable to adjust frequency bands not only individually for each subject but also for each recording site. For practical reasons, this has not yet been done.
In an empirical sense, an oscillatory component is defined by the presence of a rhythmic activity in the EEG which is manifested by a `peak' in spectral analysis. In contrast to theta, alpha as the dominant rhythmic activity, characterized by sinusoidal wave forms, clearly meets this definition. In the human EEG of young healthy adults, there are at least two other oscillatory components, the mu rhythm and the third rhythm. The mu rhythm (mu stands for motor) has an arch-shaped wave morphology, appears over the motor area and becomes suppressed (desynchronized) during motor related task demands (see the extensive work of Pfurtscheller et al., e.g., Ref. [125]). Other terms are `arcade', `comb' or `wicket' rhythm, `central', `rolandic' and `somatosensory' alpha (cf. the review in Ref. [110]p. 137f). The third rhythm which is not detectable in the scalp EEG (but e.g., by the use of epidural electrodes or magnetoencephalography [MEG]) is independent from the (posterior) alpha and mu rhythm and appears over the midtemporal region [110]. In emphasizing the fact that (in the MEG) this rhythm is best seen over the auditory cortex in the temporal lobe, Hari [51]uses the term `tau' rhythm (tau stands for temporal). There is some evidence that the tau rhythm becomes suppressed during acoustic but not visual stimulation (cf. the review in Hari et al. [52]). Because the focus of this review is on theta and alpha activity with respect to cognitive performance and memory, the mu and third rhythm will not be considered.
Despite the fact that theta does not meet the criteria for an oscillatory component in the EEG, it may still be argued that activity within the individually defined theta band reflects oscillatory processes. In a theoretical sense, the EEG can be conceived of a linear superposition of a set of different sine waves (oscillatory components). In addition, there are more specific arguments which are based on empirical evidence:
- 1.
With the help of sophisticated methods, theta peaks can be found in the human scalp EEG of young healthy adults [45].
- 2.
Theta frequency (as measured by TF) covaries with alpha frequency (as measured by IAF 34, 76).
- 3.
Theta and alpha band power are related to each other, although in a reciprocal or `opposite' way (see 2 Tonic changes and differences in the alpha and theta frequency range, 3 Interim discussion, 4 Event-related (phasic) changes in the alpha and theta bandbelow and Table 2),
- 4.
Animal research has shown that theta clearly is an oscillatory component of the hippocampal EEG which is related to memory processes (see e.g., the review by Miller [107]and Section 5below),
- 5.
Research from our laboratory indicates that theta band power increases in response to memory demands just as hippocampal theta in animals does (see Section 4.3Section 5below).
Thus, the concepts of desynchronization and synchronization will also be used for the (individually defined) theta band. In a similar way, these concepts will be applied to the different subbands of alpha. Visual inspection of `alpha waves' in the EEG may invite the misleading interpretation that there is only a single rhythm which may just vary in frequency. As the results from factor analyses have shown, there are at least two independent components of alpha activity which must be distinguished. Based on physiological and experimental evidence, many authors meanwhile assume that there is an entire population of different alpha rhythms (e.g., Refs. 6, 166, 167). Thus, it seems quite obvious to assume that during desynchronization different alpha rhythms in different subbands start to oscillate with different frequencies (for more details see Section 5).
Section snippets
Tonic changes and differences in the alpha and theta frequency range
The type of EEG changes or differences which are discussed in the following sections my be termed `tonic' in order to contrast them from `phasic' changes. Phasic (or event-related) changes in the EEG are more or less under volitional control and occur at a rapid rate, whereas tonic changes are not (or less) under volitional control and occur at a much slower rate. Phasic changes in the EEG are task and/or stimulus related. Tonic changes, on the other hand, occur over the life cycle and in
Interim discussion
If the EEG of the mature brain of young healthy adults is compared either with the developing brain, the aging brain or the brain which is affected by neurological diseases of various kinds, the conclusion is that:
(a) alpha frequency is positively related to cognitive performance, and
(b) large power in the range of the upper alpha band but small power in the theta frequency range indicate good cognitive performance.
These conclusions are based on findings which show that
(a1) alpha frequency
Event-related (phasic) changes in the alpha and theta band
Since the work of Berger it was suggested that visual (or other sensory) task demands and visual attention in particular are the primary factors that lead to a suppression of the alpha rhythm (e.g., Refs. 108, 130). In using event-related desynchronization (ERD), a method introduced originally by Pfurtscheller and Aranibar [123], recent research has revealed a much more complex picture.
A typical example of an EEG epoch which is used for measuring ERD is shown in Fig. 7. The subjects' task was
General conclusions and physiological considerations
The most important conclusion is that the amount of EEG power in the theta and alpha frequency range is indeed related to cognitive and memory performance in particular, if a double dissociation between absolute and event-related changes in alpha and theta power is taken into account. This double dissociation is characterized by the fact that during a resting state
(i) small theta power but large alpha power (particularly in the frequency range of the upper alpha band) indicates good
Acknowledgements
This research was support by the Austrian Science Fund, P-11569 and P-13047.
References (169)
- et al.
EEG alpha rhythm frequency and intelligence in normal adults
Intelligence
(1996) - et al.
Functional correlates of alphas: panel discussion of the conference `Alpha Processes in the Brain'
Int. J. Psychophysiol.
(1997) - et al.
Alpha oscillations in brain functioning: an integrative theory
Int. J. Psychophysiol.
(1997) - et al.
Discrimination of Alzheimer's disease and normal aging by EEG data
Electroencephalogr. Clin. Neurophysiol.
(1997) - et al.
Aging, brain size, and IQ
Intelligence
(1995) - et al.
Event-related desynchronization: the effects of energetic and computational demands
Electroencephalogr. Clin. Neurophysiol.
(1992) - et al.
Sleep deprivation: effect on sleep stages and EEG Power density in man
Electroencephalogr. Clin. Neurophysiol.
(1981) - et al.
Computerized EEG spectral analysis in elderly normal, demented and depressed subjects
Electroencephalogr. Clin. Neurophysiol.
(1986) - et al.
Topographic EEG changes with normal aging and SDAT
Electroencephalogr. Clin. Neurophysiol.
(1989) - et al.
EEG in children with spelling disabilities
Electroencephalogr. Clin. Neurophysiol.
(1991)