Symptom-based clusters in people with ME/CFS: an illustration of clinical variety in a cross-sectional cohort

In this cross-sectional study, members of the Dutch ME/CFS Foundation (ME/CVS Stichting; https://​mecvs.​nl/​) were invited to complete a web-based survey between January 26 and February 28 2022. Individuals who did not report being diagnosed with ME/CFS were excluded from the analyses.

Ethical approval for this study was waived by the medical ethics committee of Maastricht University because the Medical Research Involving Human Subjects Act (WMO) did not apply to this study (METC 2021-2797). Digital informed consent was obtained from all respondents at the start of the survey. Without providing informed consent, participants were unable to start the questionnaire.

Participants completed an online version of the DePaul Symptom Questionnaire version 2 (DSQ-2), which is a self-report measure of ME and CFS symptomatology, demographics, and medical, occupational and social history [19]. The DSQ-2 has demonstrated to have a strong reliability and validity [19, 20] and is able to differentiate individuals with ME/CFS from healthy controls [21] and individuals with other chronic diseases [22, 23]. Furthermore, the DSQ-2 can be used to determine whether individuals meet the criteria for the Fukuda, CCC, ME-ICC and/or IOM case definition [20]. The questionnaire was translated into Dutch using a forward–backward translation procedure.

Participants reported the frequency and severity of 79 symptoms related to the illness over the past 6 months. Frequency of symptoms was rated on a 5-point Likert scale: 0 = none of the time, 1 = a little of the time, 2 = about half the time, 3 = most of the time, and 4 = all of the time. Similarly, severity of symptoms was rated on a 5-point Likert scale: 0 = symptom not present, 1 = mild, 2 = moderate, 3 = severe, and 4 = very severe.

Symptom scores were analyzed in two ways: (1) a composite variable was created by averaging the frequency and severity scores of each symptom and multiplying it by 25; the composite score of each symptom ranged from 0 to 100 points. A higher score indicated a higher symptom burden [20]; and (2) a binary “2/2 threshold” variable was created by examining the frequency and severity scores of each symptom; participants who reported ratings of two or higher for both frequency (i.e. about half the time, most of the time, or all of the time) and severity (i.e. moderate, severe, or very severe) were considered to have the symptom [20]. For all other scores the symptom was not present.

An extra dataset of 252 people with a self-reported diagnosis of ME/CFS from the United States (US) was used to assess the reproducibility of the symptom-based clusters [24]. Participants for the US database were recruited from email requests to national foundations, posts to online support groups, research forums and social media platforms. All participants provided digital informed consent and subsequently completed an online version of the DSQ-2. Part of these DSQ-2 data were published before [24].

Data are presented as median and interquartile ranges (IQR) for continuous data and as frequencies and proportions for categorical data. Moreover, a self-organizing map (SOM) was used to visualize the clustering of the patients. The SOM method can be viewed as a non-parametric regression technique that converts multi-dimensional data spaces into lower dimensional abstractions. A SOM generates a non-linear representation of the data distribution and allows the user to identify homogenous data groups visually.

Severity and frequency scores of the 79 measured symptoms were used. Therefore each participant had 158 features. Clustering was performed on MATLAB (R2022a, MathWorks, MA, USA), following its default SOM setting, except for the number of iterations for training the SOM, which was changed to 1000 [25]. The default random number generation of MATLAB was used to initialize all competitive units of the SOM, meaning that with the same input and SOM settings, the results are always the same. Further details on the clustering method are provided in Additional file 1.

A Kruskal–Wallis test, adjusted for multiple comparisons, was used to test differences in participant characteristics between clusters. Statistical analyses were conducted using SPSS 25.0 (IBM Corporation, NY, USA). A priori, the level of significance was set at p < 0.05. The Venn diagrams were generated with Meta-Chart (https://​www.​meta-chart.​com/​venn#/​display) and intersection plots were made using the R package ‘UpSetR’ [26]. Finally, the trained SOM was applied to the data from the validation sample.

The link to the online questionnaire was send to 1392 members of the Dutch ME/CFS Foundation, of which 367 completed the questionnaire (response rate: 26%). Data from thirty participants were excluded, as they stated to not have been diagnosed with ME/CFS. So, data from 337 participants were used for the analyses (Table 1). In general, participants were mostly middle-aged women [82% female; median (IQR) age: 55 (44–63)] with a normal body mass index [24 (21–28) kg/m²]. Fifty-five percent of the participants were married or living with a partner. The majority of the participants had a medium or high education level and about two-thirds of them were incapacitated for work. Almost 90% of the participants fulfilled the Fukuda case definition, compared to 80%, 59% and 39% fulfilling the IOM, CCC and ME-ICC case definitions, respectively. More than a quarter of the participants met the criteria for all four different case definitions, whilst 5% of the participants met none of the abovementioned case definitions (Table 1, Fig. 1a).

The vast majority of the participants were experiencing fatigue (90.7% fulfilled the 2/2 threshold). Furthermore, PEM, sleep-related problems, neurocognitive problems and pain were frequently reported, whilst on average autonomic, neuroendocrine, and immune symptoms were less prevalent (Table 2). Participants reported a median (IQR) of 27 (19–37) symptoms (using the 2/2 threshold, Additional file 2).

Forty-five clusters were identified, of which 13 clusters included ≥ 10 patients (Fig. 2a, b). In general, key features of these 13 clusters can be summarized as follows:

Participants in Clusters 28 and 37 reported a significantly lower number of symptoms compared to participants in Clusters 4, 7, 9, 24, 26, 31 and 36, whilst the number of symptoms in Clusters 36 was significantly higher compared to all other clusters, except for Clusters 7, 9 and 26 (p < 0.05; Table 3).

Please see Additional file 3 for symptom scores of the different clusters. Please see Additional file 4 for the symptom scores of the remaining 42 clusters with < 10 patients.

Participant characteristics of the 13 largest symptom-based clusters are displayed in Table 3. On average, participants in cluster 4 were significantly younger compared to participants in clusters 7, 19, 24, 28, 37 and 40. Cluster 19 had a significantly higher proportion of males compared to Clusters 28. Prevalence of work disability was significantly lower in Cluster 37 compared to Clusters 4, 7, 24, 26 and 36. Please see Additional file 5 for the characteristics of the remaining 42 clusters with < 10 patients.

Distribution across the different case definitions and number of case definitions that were met were significantly different between the 13 largest clusters. Generally, the proportion of patients fulfilling the different case definitions was highest in Cluster 9 and lowest in Cluster 19, 28 and 37. None of the participants in Cluster 37 fulfilled the CCC and ME-ICC case definition, and also the proportion of patients fulfilling the IOM case definition was lowest in Cluster 37. The proportion of patients fulfilling the Fukuda CFS Criteria was lowest in Cluster 36. In addition, Clusters 9, 26 and 36 had the highest proportion of participants fulfilling the ME-ICC case definition.

Participant characteristics of the US database are listed in Table 1. In general, these participants were slightly younger, were more often widower or divorced, had a higher education level and were less often on disability compared to the Dutch participants (all p < 0.05). The proportion of patients fulfilling the different case definitions and the total number of case definitions met was significantly higher in the US database compared to the Dutch database. Furthermore, US participants generally experienced a higher symptom burden compared to the Dutch participants (Table 2). Applying the trained SOM to the data of the validation sample, resulted in a similar symptom pattern compared the Dutch dataset (Fig. 2c).

A clear strength of this study is the use of two datasets with a considerable amount of participants with ME/CFS to identify and apply the symptom-based clusters. This supports the external validity of our findings. Furthermore, using the SOM approach, we were able to cluster a large dataset and visualize it on a two-dimensional map, in which similar datapoints are clustered into the same group or nearby groups.

This study has the following methodological limitations. First, the possibility of selection bias is present, as it is reasonable to assume that participants with a higher symptom burden are more likely to complete the questionnaire. On the other hand, patients with cognitive or concentration related problems are less likely to participate. To account for concentration related problems, participants were allowed to take breaks while completing the online questionnaire. Second, similar to earlier studies, the used populations predominantly consisted of females [29, 30], though it has been recognized that prevalence of ME/CFS is higher in women compared to men [31]. Furthermore, almost all participants were white, whilst however the prevalence of ME/CFS has suggested to be higher in ethnic minority populations [32, 33]. Third, the DSQ-2 only captures the frequency and severity of symptoms, but doesn’t account for within-day and between-day variation in symptoms. Future studies should consider the use of ecological momentary assessment, which involves repeated measurements of the participant’s symptoms, behavior and context in vivo and in real time [34]. Fourth, clusters were based on self-reported symptoms and did not include clinical variables, such as inflammatory markers, physical functioning, anxiety, depression, common comorbidities and/or current treatment. Moreover, the online survey did not allow us to identify whether symptom-based clusters were associated with relevant patient-reported outcomes, including functional status and quality of life.