This study was part of a larger effort to evaluate the occurrence of, and risk factors for, CFS and unwellness in the 18- to 59-year-old population of the state of Georgia. Figure 1 summarizes the sample, and details are available elsewhere [17]. Briefly, between September 2004 and July 2005, we used random digit dialing to conduct a household screening interview with a household informant in three Georgia populations (metropolitan, urban, and rural). A household informant described demographics and health status of household members ages 18 to 59. That initial interview enumerated 19,807 adult residents and screened for unwellness among household members for whom at least one CFS symptom was reported (fatigue, impaired cognition, un-refreshing sleep, muscle or joint pain). Well residents had none of these symptoms for ≥ one month. The screening interview identified 10,834 (55%) well people; 5,122 (26%) people who were unwell for at least a month but not fatigued; and 3,851 (19%) people who were unwell and fatigued for at least a month. We then conducted detailed telephone interviews with all those identified as unwell with fatigue, a random selection of those who were unwell but without fatigue, and a random sample of well people (see Figure 1). Based on their responses to the detailed telephone interview, we classified participants as CFS-like if they met criteria of the 1994 CFS case definition [5]; as chronically unwell if they exhibited some but not all CFS symptoms; and as well if they reported no such symptoms. Finally, we invited all 469 people classified as CFS-like; 641 well people (matched to the CFS-like people by sex, race/ethnicity, age, and geographic stratum); and a similar number (n = 505) of randomly selected unwell people for a one-day clinical evaluation. Overall, 48.5% completed the clinical evaluation.

To identify medical conditions considered exclusionary for CFS [5,7], the clinical evaluation included a standardized past medical history, a review of systems, a standardized physical examination, and routine laboratory testing of blood and urine [5,7]. To identify psychiatric conditions considered exclusionary for CFS, licensed and specifically trained psychiatric interviewers administered the Structured Clinical Interview for DSM-IV (SCID) [18]. Exclusionary psychiatric disorders included current melancholic major depression, any psychotic condition, bipolar affective disorder, active substance abuse/dependence, anorexia or bulimia nervosa [5,7]. Subjects with current non-melancholic major depressive episode or a past history of major depressive disorder were not excluded. The study identified exclusionary medical or psychiatric conditions in 280 (36%) of those evaluated clinically. They were excluded from further analyses, as were two others with incomplete data, leaving a total sample of 501 subjects.

We also diagnosed CFS based on the clinical evaluation according to criteria of the 1994 case definition [5] and as recommended by the International CFS Study Group [7], which is standard in CDC studies of CFS. Thus, we evaluated functional impairment by means of the Medical Outcomes Short-Form Health Survey (SF-36) [19]. We used the Multidimensional Fatigue Inventory (MFI-20) [20] to measure characteristics of fatigue, and we utilized the CDC CFS Symptom Inventory to document occurrence, frequency, and severity of the defining symptoms [21]. Subjects who had ≥ four case-defining symptoms, who exceeded the Symptom Inventory cut-off score, and who met CFS cut-off scores on the SF-36 and the MFI-20 were considered to have CFS (n = 113 participants). Those who met at least one, but not all CFS criteria, comprised the ISF group (n = 264), and those who met none of the cut-off criteria comprised the Well group (n = 124). In consideration of inadequate sample size for detecting the heterogeneity in both sexes, and for replication of the Wichita study (which only studied women), we opted to only study women. This left 386 women in the analysis for this paper.

Although many of the 38 variables used in the Wichita study were the same (see Table 1), some, such as nocturnal polysomnography, were not measured in the Georgia study. In all, 26 clinical and laboratory variables were the same or very similar to the 38 variables used in the Wichita study [13]. When a variable was not available, we used proxy measures to estimate the effect of the missing variable (see Table 1). Cortisol index was the sum of four serum cortisol concentrations taken regularly over 24 hours minus the baseline concentration.

In order to externally validate the subgroups, we used the Short-Form Health Survey measure of functioning (SF-36) [19] and the Multidimensional Fatigue Inventory [20], neither of which had been used to define the classes.

We used Latent GOLD^® (LatentGOLD 4.1.3, Statistical Innovations, MA, USA) software to perform latent class analyses (LCA) on all variables and all subjects. Non-categorical variables were dichotomized by median split, and we filled missing (null) values (0.4% missing data values) with an intermediate categorical value. LCA is a statistical method for finding subtypes of similar subjects (latent classes) using multiple, defining, and categorical variables. LCA searches for an optimal class assignment for each subject so that most of the variation in the entire dataset can be explained by the subject's class assignment alone. LCA posits the existence of statistically uncorrelated measurements within each of the discovered classes. One therefore seeks measurements that most fully define the variability in the dataset, but are uncorrelated with each other. LCA and other clustering techniques seek to find loci in the multidimensional space of measurements (clinical and biological) where subjects cluster together. The subjects that are close to each other belong to the same class. Finding the locus for each class requires fitting one parameter for each dimension (measurement), and specifying a locus in an n-dimensional space requiring n parameters. To determine the optimal number of latent classes, we utilized the Akaike Information Criterion (with 3 as penalizing factor) (AIC3). AIC3 is defined by the following equation:

where L is the likelihood and npar is the number of estimated parameters.

Recent results [22,23] suggest AIC3 is a better criterion than Bayesian or AIC for LCA models. We utilized AIC3 for the Wichita study and also utilize it here. The optimal number of classes is determined through locating a minimum in the value of AIC3 for solutions with different numbers of clusters. To validate the different classes, we compared independent variables across the empirically defined classes using Kruskal-Wallis or Mann-Whitney U tests for interval variables. Variables tested for validity included the Short-Form Health Survey (SF-36) and the Multidimensional Fatigue Inventory. A final test of validation was used to compare the empirically defined classes against a standardized application of impairment, fatigue, and symptom criteria of the 1994 CFS case definition [24] as recommended by the International CFS Study Group [7] and the screening criteria of "unwell and not fatigued" and "well."

The median values on the variables originally omitted from deriving optimal class solutions are shown for each of the five classes in Table 3. With all five classes included in the analysis, nonparametric ANOVA produced highly significant overall tests indicative of substantial between-class differences for all variables (see Table 4). We used Mann-Whitney U follow-up tests for fatigue and disability scores to compare each of the three ill classes against the two well classes (Classes 4 and 5). This confirmed significant differences across all the dimensions of disability and fatigue between Classes 1 and 2 and the well Classes 4 and 5 (all P values < 0.001). Comparisons among the less symptomatic but still ill Class 3 and both of the well classes still produced highly significant differences for physical functioning, and for general, physical, and mental fatigue (all P values < 0.008). Similarly, there were significant differences in activity and motivation (all P values < 0.009). Reported levels of mental and social functioning were substantively different only between Class 3 and Class 5 (P values < 0.003).

Significant between-class differences also distinguished the three ill/fatigued classes in all disability and fatigue categories (Table 4). Follow-up tests showed no significant differences between Classes 1 and 2 on the SF-36 dimensions [19] of mental functioning and the Multidimensional Fatigue Inventory subscales [20] of general fatigue, mental fatigue, and reduction in motivation (all P values > 0.05). However, Class 2 reported significantly lower physical (P < 0.000) and social functioning (P < 0.01) and higher levels of physical fatigue (P < 0.007) and activity reduction (P < 0.009) than Class 1. Comparisons of SF-36 and fatigue scores between Classes 1 and 3 revealed no statistically significant difference in SF-36 physical functioning, but differences on the dimensions of mental and social functioning (both P values < 0.000). Additional differences between these two classes were evident on general fatigue, mental fatigue, and reduction in motivation (all P values < 0.000), and to a lesser degree for physical fatigue (P < 0.01) and activity reduction (P < 0.04). Comparisons between the most symptomatic Class 2 and the least symptomatic Class 3 of the ill classes were all significantly different at the P < 0.000 level.

Table 5 shows the distribution of criterion-based diagnoses: CFS, unwell but not fatigued, and well individuals. People diagnosed as CFS fell almost exclusively into Classes 1 and 2. Those who were unwell but not fatigued were fairly evenly spread across all classes, while the well were most commonly found in Classes 4 and 5, with a small minority in Class 3.

Table 6 compares and contrasts the four class solutions across the two studies in Georgia and Wichita. There was a close similarity in the content of all four classes. Table 7 compares and contrasts the five class solutions across the same two studies. Four out of the five classes were very similar in their content.

How do these results compare with the Wichita classes? First, the main variables that differentiated the classes were similar: obesity, metabolic strain, multiplicity of physical symptoms, sleep problems, and depression. Secondly, the four-class solutions were very similar across the two studies, and four out of the five class solutions were also similar.

The dissimilarities include empirically describing a five-class solution in Georgia compared to a four-class solution in Wichita. The absence of measures of sympathetic nervous system arousal, such as nocturnal heart-rate variability and polysomnography, used in the Wichita study, may explain the lack of further similarities between the two samples. Similarly, the inability to differentiate a sixth class, similar to that reported in the Wichita sample, was probably related to the absence of the sex steroid hormone assays in the Georgian sample because the sixth class was defined primarily by being post-menopausal [12]. The SF-36 physical function subscale scores were higher than we expected, but the scores were similar to those of the relevant subjects from the Wichita study and probably reflect the sample being derived from the population rather than a clinic. Previous studies of CFS have also used a previous version (version 1 rather than 2) of the SF-36 scale, with a slightly different stem question, which also may have affected the data.

The multiplicity of symptoms in some classes supports the two previous studies differentiating CFS on the basis of a minority having features of somatisation [8,9]. The importance of depression reflects the differentiation found on the basis of mood disturbance in the largest study of heterogeneity using the Swedish twin registry, which found two main subgroups, one of which was differentiated primarily by comorbid mood disorders [25]. Kerr and colleagues' finding of seven subgroups was based on secondary and tertiary care samples and was determined by gene expression alone [16]. Apart from the Wichita study, for which this is a replication, no other study has defined subgroups using both clinical and laboratory data, so no other comparisons are possible.

The most important limitation was the absence of some variables that had been used in the Wichita study. These primarily included objective measures of sleep, such as polysomnography, and autonomic nervous system tests such as heart rate variability. Our employment of proxy variables was unsuccessful because they rarely played a differentiating role in the analyses. We were careful to otherwise keep the variables and their preliminary processing, as well as the analyses, the same. In view of the discrepancy in important variables, it is remarkable that so many similarities were found. The addition of other variables, which neither study measured, such as actigraphy and other psychiatric comorbidities (e.g., anxiety), might have further enhanced the differentiation of classes [26]. As in all population surveys, there may have been a response bias of which we were not aware. Finally, this replication was based on American women, as in the Wichita study, so that these data cannot necessarily be generalized across other nations or to men.

What are the clinical implications of this work? This study's support for the heterogeneity of CFS, found in an independent population-derived sample, adds to the weight of evidence for the heterogeneous nature of CFS. We do not yet know whether this heterogeneity is important in determining response to different treatments, but future research studies should now examine for moderators of outcome that include obesity, metabolic syndrome, sleep problems, depression, and having multiple symptoms.

As in Wichita, these data suggest that the current research criteria for CFS, while useful in providing reliable findings across studies, do not constitute a homogeneous group of patients. The broadening of the concept of CFS to include patients with fewer symptoms but similar disability is supported by this replication [6,7].

What are the research implications of this work? Future etiological work needs to take into account the heterogeneity of CFS, using the determining variables, such as obesity, multiple symptoms, and depression, to stratify samples. For instance, obesity leads to fatigue in multiple ways, including sleep disturbance, metabolic strain, increased inflammatory markers, and deconditioning [12,27,28]. This heterogeneity may well explain the dearth of replicated etiological findings in CFS. Addressing heterogeneity up front may help to determine the different etiologies and pathophysiologies that likely determine the different illnesses that make up CFS.