Background
Fatigue is one of the most common presenting complaints, accounting for a 10-25% prevalence of patients presenting to primary care physicians (PCP) [
1]. The extensive differential diagnosis of fatigue encompasses a wide spectrum of illnesses including, but not limited to endocrine disorders, infections, cancer, medication side effects, sleep disorders, seizures, autoimmune diseases, obesity, drug abuse, malingering, and depression [
2]. Fortunately, most of these illnesses have characteristic clinical presentations often with confirmatory laboratory tests.
Yet there remain significantly fatigued patients where no underlying diagnosis can be securely established. In the past, such patients were often dismissed as having some form of uncertain psychiatric disorder-typically depression with symptoms of somatization. However, within this 'unclassifiable' but severely fatigued patient population a subset stood out with normal pre-morbid personalities and whose pre-morbid lives were successful and fulfilling. These patients, however, had suddenly become unusually fatigued after an undetermined illness and for whom the subsequent disabling weakness and fatigue endured for more than six months (often years) beyond the resolution of the initial illness. Some, but not all, patients would report intermittent lymphadenopathy and/or low grade fever often with corresponding worsening of their fatigue. Yet, no clear etiology could be found. The term Chronic Fatigue Syndrome (CFS) came to be applied to this group where a suspicion of organic etiology persisted but could not be confirmed [
2,
3].
Since common psychiatric disorders, particularly depression, often cause fatigue and since psychiatric diagnoses may be difficult to objectively and reliably confirm, many continued to reasonably wonder about the role of an as of yet identified form of depression as the cause of CFS. However, it was found that many patients with CFS suffer from co-existing psychiatric disorders only
after becoming ill with CFS. Moreover, in 30-50% of patients no co-existing psychiatric disorders [
4,
5] can be demonstrated. In addition, a carefully controlled trial of fluoxetine in patients with CFS failed to improve fatigue, even in those patients with a concomitant major depression [
6].
To better identify this perplexing patient population, the U.S. Centers for Disease Control (CDC) convened a group of experts to establish a set of strict diagnostic criteria for CFS. The resultant criteria have become known as the CDC or Fukuda criteria [
3]. These criteria, available as a multi-page evaluation form, serve investigators and clinicians studying CFS to assure that their patient populations are well identified and comparable across studies. CFS is, therefore, not a synonym for prolonged, disabling fatigue although the distinction may be difficult upon initial evaluation. In this paper we use the term CFS to mean CDC-defined CFS.
CFS-which constitutes 0.5-2.5% of primary care referrals and 10-15% of tertiary care referrals for fatigue [
1] -remains without confirmatory laboratory tests and can be difficult to distinguish from depression. Between 1 and 8 in 1000 U.S. adults meet the CDC criteria [
7]. The CDC estimates that cost to the U.S. economy from lost productivity alone (not including medical care costs) is $9 billion annually [
8].
There exists published evidence that CFS may have its underpinnings in organic disease especially within the central nervous system (CNS), although not all studies have found such abnormalities. Studies of the CNS in CFS have included psychometric assessment of cognition [
9,
10], magnetic resonance imaging [
11‐
13], functional MRI [
14,
15],
in vivo MR spectroscopy [
16,
17], single-photon emission computed tomography [
18], positron emission tomography [
19], neuroendocrine studies of hypothalamic function [
20‐
22], and studies of the autonomic nervous system [
23‐
25].
A link with infection and CFS also has been reported following infection with Epstein-Barr virus, Ross River virus,
Coxiella burnetii [
26],
Borrelia burgdorferi [
27], parvovirus B19 [
28], human herpesvirus-6 [
29], and enteroviruses [
30]. Novel retroviruses may also be involved [
31,
32] but that possibility has been challenged [
33]. All these infectious agents have the potential to be CNS pathogens. The evidence of neurologic involvement in CFS, and the possible role of infectious agents in triggering and perpetuating CFS, is summarized in a recent review [
34].
Symptoms suggesting the possibility of subtle encephalitis in CFS, along with the documented association of CFS with several neurotropic infectious agents, caused us to examine the role of electroencephalographic (EEG) studies in this illness. However, simple visual inspection of EEG has rarely provided valuable information in CFS, aside from allowing exclusion of epilepsy and classic encephalopathy. A study utilizing EEG Spectral Analysis [
35] reported no significant differences of spectral power in any EEG frequency bands during sleep between subjects with CFS and their non-fatigued co-twins. Only studies requiring stressful conditions such as repetitive muscular exercise [
36] and sleep deprivation [
37] have documented EEG spectral difference in CFS.
Accordingly, we undertook an exploration of spectral coherence, a more complex computational derivative of EEG spectral data, which estimates connectivity between brain regions [
38‐
40]. We hypothesized that results would, first, serve to confirm a consistent pattern of brain difference in CFS and, second, provide estimates of the potential for an EEG based diagnostic test for CFS.
Results
Identification and Selection of Spectral Coherence Variables
Variance distribution among the resulting coherence factors was favorable: 2014 factors described over 99%, 302 described 90.03%, 37 described 50.32%, 7 described 26.01% and 1 described 8.25% of the total variance. The first 40 factors-accounting for 55.64% of total variance-were chosen for analysis, exceeding Bartlett's recommendation [
72] and resulting in a conservative sample size to variable ratio of 235:40 or 6:1 [
73] for the initial DFA described below.
Discriminating Groups Using Spectral Coherence Variables
The primary discriminant analysis was based on the 197 unmedicated female controls and 38 unmedicated female CFS patients. Female subjects were chosen because in most case series and epidemiologic studies of CFS, females outnumber males [
7].
When all 40 coherence factors were forced to enter the DFA, there was a highly significant (p < 0.0004) group differentiation by Wilks' Lambda, with Rao's approximation. The unmedicated female CFS patients were identified with 89.5% accuracy and the female controls with comparable 92.4% accuracy. Age did not significantly differ between these two groups. The statistically significant result, with all 40 factors as variables forced to enter, establishes that these two groups differ on the basis of variables generated from EEG based coherence data.
Stepwise DFA was then utilized to identify a factor subset that best described the group difference. Ten factors (Figure
2, Table
2) entered the model resulting in a highly significant discrimination (p < .001) and equivalent classification success rate: unmedicated female controls 89.85%; unmedicated females with CFS 86.8%. Loadings of the 10 best factors (Table
2) determined to be useful in subsequent group discriminations are topographically displayed in Figure
2.
Table 2
Coherence Loadings on 10 Best Coherence Factors
1 | +0.91 | 2-6 | OZ ← → FT9, F7, FP1 |
| | 2-6 | O1 ← → FT9, F7, FP1 |
| | 2-6 | O2 ← → FT9, F7, FP1 |
2 | +0.82 | 24-28 | OZ ← → T7, FC5, F3, C3 |
3 | -0.80 | 20-26 | T8 ← → FC6, CP6 |
19 | -0.64 | 18-28 | F3 ← → CP5, FC5, FC1, CP2 |
27 | +0.61 | 6 | T7 ← → P7, FC6, T8, CP6 |
| | 6 | T8 ← → FC5, T7, CP5, P7 |
21 | -0.61 | 4-10 | C3← → FC5, FC2, C4, T8, F8 |
| | 4-10 | C4 ← → C3, FC5, F8 |
| | 14-26 | C3 ← → C4, T8, F8 |
| | 14-26 | C4 ← → C3, T7 |
24 | +0.58 | 8 | F4 ← → F3, P7, TP9, CP2, P8, TP10 |
28 | +0.57 | 2-8 | T7 ← → FC1, C3, CP1 |
37 | +0.55 | 14-24 | P4 ← → CZ, CP1, O2, P8 |
20 | -0.35 | 8-12 | FC1← → CP1, F7, FP1, FP2 |
The results of the 10 jackknifing trials are shown in Table
3. The average success for the ten trials is reported for the control (87.14%) and CFS females (86.2%). Each of these ten iterations generates a unique canonical discriminant variable for each test set member on the basis of the corresponding training set data. As a separate measure of classification success a 2-group analysis of variance (ANOVA) is performed for the discriminant variable on test set subjects (BMDP -7D). All of the 10 iterations reached significance, eight at or below the p < 0.0003 level, one at the p < 0.006 level and one at the p < 0.02 level.
Table 3
Recursive Jackknifing by Leaving 20% Out: Test Set Classification Accuracy
1 | 35/41 | 85.36 | 8/9 | 88.89 | 1,14 | 38.09 | 0.0000 |
2 | 34/38 | 89.47 | 5/5 | 100.00 | 1,5 | 20.42 | 0.0063 |
3 | 32/39 | 82.05 | 9/10 | 90.00 | 1,19 | 39.66 | 0.0000 |
4 | 36/41 | 87.80 | 8/9 | 88.89 | 1,11 | 41.38 | 0.0000 |
5 | 37/41 | 90.24 | 5/6 | 83.33 | 1,6 | 9.17 | 0.0232 |
6 | 35/39 | 89.74 | 8/10 | 80.00 | 1,14 | 22.51 | 0.0003 |
7 | 33/43 | 76.74 | 8/9 | 88.89 | 1,14 | 29.51 | 0.0001 |
8 | 41/47 | 87.23 | 7/9 | 77.78 | 1,11 | 29.89 | 0.0002 |
9 | 40/44 | 90.90 | 11/14 | 78.57 | 1,28 | 51.75 | 0.0000 |
10 | 36/39 | 92.31 | 6/7 | 85.71 | 1,10 | 43.47 | 0.0001 |
Mean
| |
87.14
| |
86.21
| | | |
By both classification success and ANOVA, results were positive for use of spectral coherence data in prospective classification.
Applying the Discriminant Function to Other Groups
The 10-factor discriminant function derived from the unmedicated female subjects was then tested on the other patient groups. Of note, 8 of the 9 (88.9%) unmedicated CFS males, whose data were not included in formation of the discriminant formation, were correctly classified.
The discriminant function was applied to male and female CFS subjects who were taking psychoactive medications. Although it performed considerably better than chance, the discriminant performed less well than it had with unmedicated subjects: 14/18 (77.8%) of medicated female CFS patients and 3/5 (60%) of medicated male CFS patients were accurately classified.
For patients with unspecified fatigue whether medicated or unmedicated, 46.6% were assigned to the CFS classification. As the true diagnosis of these subjects is not known, accuracy of the classification cannot be inferred.
Finally, when the discriminant function was applied to all four subgroups of the 24 patients with major depression, none of the depressed patients were falsely classified as having CFS.
Characteristics of Coherence Variable Differences between CFS and Normal Subjects
There was no clear predominant side (right vs. left) or EEG spectral band involved in the 10 factors that were the best discriminators. However, there were clear differences in the brain regions involved in the ten most discriminating coherence factors, as follows: Temporal region (9/10), central (8/10), frontal (5/10), occipital (3/10), and parietal (1/10) region. (Figure
2)
Discussion
The first goal of this study was to explore meaningful reduction, by principal components analysis (PCA), of a large data set of artifact-free EEG spectral coherence data created from an adult population containing healthy controls and patients with CFS, major depression, and unspecified severe fatigue. Coherence is taken to represent the degree of functional connectivity or coupling between two different brain regions at a chosen frequency.
The
second goal was to explore the utility of the PCA-reduced data set in differentiating CFS patients from normal subjects without falsely classifying depressed patients as having CFS. Many studies have found evidence of nervous system involvement in CFS, but no large, controlled investigations of the value of EEG spectral coherence in patients with CFS had been reported. Spectral coherence has proven useful in conditions where standard EEG is seldom found to be diagnostic [
59,
71,
74,
75].
First goal, creation of artifact free coherence factors by PCA
Utilizing the full subject population (Table
1, n = 632) we were successful in reducing the initial 7936 coherence variables per subject to 40 orthogonal (uncorrelated) factors per subject which described 55.6% of the total, initial variance. In other words, PCA condensed over half the information (variance) contained in the initial 7936 variables into just 40 new variables (outcome factors). One benefit of this almost 7936:40 or 200 fold reduction in data dimensionality over the entire population is a parallel reduction in the likelihood for capitalization on chance of the sort that may occur during subsequent statistical analyses when they involve large numbers of variables [
48]. An additional benefit to this 'hands-off' data reduction is that it requires no advance or
a priori coherence variable selection by the investigators, eliminating any possible variable selection bias. Bartels refers to this as allowing the intrinsic data structure of the population to select variables [
45].
In utilizing this PCA based approach, it is important to include all subjects in the initial PCA, even including subjects with related but not completely defined clinical diagnoses-in our case medicated patients and generally fatigued patients with incomplete diagnoses. Among-subject variance within the population is responsible for factor formation. For instance, had factor formation been limited to healthy normal control subjects exclusively, the degree of variance introduced by fatigue, depression and medications would have, therefore, been absent and factors potentially important to group separation might never have been formed.
Finally, the data underwent an initial multiphase artifact control process (see Methods) performed across the entire population. It is highly unlikely that the final, processed coherence data contained significant eye movement or muscle contamination. Indeed prior to PCA, the coherence data were processed so as to be uncorrelated with six classic measures of eye and muscle artifact. Thus it is unlikely that our study findings reflect artifactual group differences.
Finally, subject selection for the primary study groups (healthy controls, CFS, depression) was rigorous and performed by clinical experts in their fields on the basis of standardized, published criteria. This will facilitate replication including sample selection for future studies here and/or elsewhere.
Second goal, differentiating CFS patients from healthy controls
Our study findings indicate that EEG spectral coherence data, recorded in the waking eyes closed state, differ significantly between healthy control female subjects and otherwise healthy female patients with CDC-defined CFS. Our 40 coherence factors, significantly separated these two index subject groups at p < 0.001. This fundamental finding indicates that CFS patients manifest patterns of functional brain coupling that differ from those of normal controls. Such a difference of CFS brain physiology may help explain known differences in cognition, memory, sleep, and affect that afflict CFS patients (see Background).
We also found that a small subset of as few as 10 coherence factors were able to accurately identify (by stepwise discriminant analysis) these same unmedicated female subjects (CFS 86.8% accuracy, control 89.8% accuracy). When the rules generated by this analysis on unmedicated females were prospectively applied to unmedicated CFS males and healthy control males who were not involved in the discriminant function creation, true prospective classification accuracy remained high (CFS 88.9%, control 82.4%). In addition, when the classification rules were applied to the entire depressed population, none were falsely, prospectively, classified as having CFS.
Jackknifed classification techniques, employed to provide estimates for the prospective success rate for application of the discriminant rules to new sets of unmedicated female subjects (CFS and normal), was successful. By a re-iterative leaving 20% out processes, accuracy for controls was 87.1% and for CFS was 86.2%, (Table
3). Thus the discriminant should prove effective on entirely new samples. However, that hypothesis must be tested on a large, new set of patients with CFS and comparison groups (healthy and with other fatiguing illnesses) to assure the accuracy and utility of EEG spectral coherence as a diagnostic aid.
Speculations
The less than 100% accuracy of our spectral coherence based classification function could reflect a deficiency in the CDC criteria for CFS, and/or a deficiency in the coherence-based discriminant itself, and/or unexplored physiological variability even within carefully CDC-defined CFS. For example, multiple etiologic agents have been identified as potential triggers of the CFS phenotype [
26], each with the potential for a slightly differing impact upon the central nervous system (CNS) and, hence, on EEG spectral coherence. The possibility of sub-grouping [
76] CFS on the basis of coherence and other objective CNS measures (e.g., MRI, SPECT/PET, neuropsychology) may be a fruitful area for further exploration. Subgrouping could result in a broader set of objectively derived CNS measures from neurophysiology and other neuroimaging techniques that might eventually become the diagnostic 'gold standard' for CFS.
When applied to patients with CFS who were taking psychoactive medications at the time of testing, the 10-factor model was less accurate (females, 77.8% accuracy; males, 60.0% accuracy). Since psychoactive medications directly affect the brain, the organ being examined by EEG, it is possible that these medications may modify EEG measures such that their accuracy is compromised. Alternatively, these medications may have had a therapeutic clinical effect on brain function (connectivity), thus causing some CFS patients to electrophysiologically resemble normal controls. Supporting this hypothesis is the observation that some patients were tested while on psychoactive medications because they refused to discontinue them being convinced from past experiences that this might worsen their clinical condition. Thus another fruitful area for further exploration is to determine if EEG spectral coherence is a useful index measure in assessing medication treatment response.
Given a lack of detailed clinical information, it is not possible to determine classification accuracy within our Unspecified Fatigue population. When the 10 coherence factor discriminant is applied to this group 46.6% are classified as CFS. This is broadly consistent with the published estimate that the prevalence of true CFS among patients seeking care from tertiary specialists for prolonged fatigue can be as high as 35% [
1].
The finding of bilateral temporal lobe involvement in 9 of 10 factors is of potential clinical significance. The 10 coherence factors did not collectively localize to any other single brain region. This greater temporal lobe involvement is consistent with the global memory impairment in CFS reported by Marcel [
9] and Daly [
77]. It is also interesting that one neurotropic virus associated with CFS, human herpesvirus-6, appears to selectively affect the temporal lobes and has recently been associated with temporal lobe seizure disorders [
78‐
80].
Future plans
Our immediate plans call for enlarging our population to prospectively test and refine current findings. This will primarily involve recruiting additional patients with depression and non-CFS prolonged fatigue as well as additional patients with CDC-defined CFS-especially males. All patients will have equivalent evaluations: clinical and behavioral as well and neurophysiological. We plan to evaluate a population of CFS patients before and after beginning medications. We also hope to develop specific classification rules to separate four diagnostic groups: CFS, non-CFS prolonged fatigue, depression, and healthy controls. We plan to search for CFS-gender interactions. All this will require substantially larger populations than now available to us. Finally, within the CFS population we will employ cluster analysis, as successfully applied by Montironi and Bartels [
76] in another research area, to search for consistent CFS subpopulations.
Acknowledgements and Funding
The authors thank registered EEG technologists Ellen Belles, Jack Connolly, Vincent DaForno, Herman Edwards, Susan Katz, Sheryl Manganaro, Marianne McGaffigan, and Adele Mirabella for the quality of their work and for their consistent efforts over the years. The authors also wish to thank Christopher Suarez for administrative and technical support for manuscript preparation and submission. These individuals so acknowledged performed their roles as part of regular duties and were not additionally compensated for their contribution.
The authors specially thank Marilyn Albert, PhD (Johns Hopkins University) for contribution of normal healthy control subjects, Allan Shatzberg, MD (Stanford University) for contribution of the patients with major depression, and Kenneth Jones, PhD (Brandeis University), author of the specialized PCA software used herein. None received compensation for their contribution to this manuscript.
This work was supported in part by National Institutes on Aging program project PO1AG049853 to M. Albert; National Institute of Neurological Diseases and Stroke grant NS1367 to F.H. Duffy; National Institute of Child Health and Human Development grant HD13420 to F.H. Duffy, the Mental Retardation Program Project P30HD18655; and by a gift from the De Young Foundation.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
The authors' contributions included the following: study concept and design, FD, GM and AK; acquisition of patients, AK and GC; acquisition of the patient data, AK, MM, FD, GM and GC; preparation of neurophysiologic data, FD and MM; analysis and interpretation of the data, FD, GM and AK; and statistical analysis; drafting and revision of the manuscript, FD, GM, AK and MM. FD had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. All authors read and approved the final manuscript.