Cross-sectional CFS study links fatigue to six symptom clusters

In a cross-sectional study of 214 adults with chronic fatigue syndrome, fatigue was the most prevalent symptom and had the greatest estimated influence within a network linking physical, cognitive, emotional, sleep, gastrointestinal, and self-image symptoms. Researchers at the First Affiliated Hospital of Zhejiang Chinese Medical University in Hangzhou, China, combined principal component analysis with symptom network modeling to identify six clusters and estimate which symptoms were most strongly connected to the overall symptom system and to other clusters. Because the study was observational, conducted at one outpatient clinic and based on self-reported symptoms, its network findings describe associations in this sample rather than causes, prognosis or proven treatment targets.

Centers for Disease Control and Prevention criteria for chronic fatigue syndrome, and each had at least six months of unexplained fatigue accompanied by the required additional symptoms. Recruitment took place between September 18, 2025, and April 2026 at a Traditional Chinese Medicine internal medicine outpatient clinic. Of the 225 questionnaires distributed, 214 were valid, producing a response rate of 95.11%.

The sample was predominantly female, with 142 women, or 66.36%, and 72 men. The largest age group comprised 93 patients aged 31 to 40 years, or 43.46%. College or undergraduate education was reported by 146 participants, and 37 had a master’s degree or higher.

Symptom burden was extensive. Fatigue was reported by every participant, difficulty concentrating by 194 patients, or 90.7%, and feeling drowsy by 186, or 86.9%. Nineteen symptoms occurred in more than half of the sample.

Shortness of breath had the highest mean severity score in the study’s main table, at 2.31, followed by feeling nervous at 2.28; fatigue had a mean severity score of 1.94. Researchers recorded symptoms as present or absent and, when present, assessed their frequency, severity and distress over the preceding seven days. They used a 35-item adapted Memorial Symptom Assessment Scale, adding bradyphrenia, chills, noise sensitivity and palpitations while removing an item about problems with sexual interest or activity because prospective participants frequently declined to answer it.

For cluster identification, the researchers first performed exploratory factor analysis with principal-axis factoring and Promax rotation, using a permutation-based parallel analysis to determine the number of factors. They then applied principal component analysis with varimax rotation to symptoms present in at least 35% of participants. Components were retained when they had an eigenvalue of at least one, contained at least two symptoms and produced loadings of at least 0.40.

Dry mouth was excluded after its loading fell below that threshold in the relevant analysis. The final model identified six components with eigenvalues greater than one, together explaining 63.114% of the variance. The Kaiser–Meyer–Olkin measure was 0.787, while Bartlett’s test of sphericity gave χ² = 1,625.864 with P < 0.001, indicating that the data were suitable for factor analysis.

The resulting clusters were emotional-sleep, fatigue, physical, self-image, gastrointestinal and neurological symptoms. The fatigue cluster contained difficulty concentrating, bradyphrenia, fatigue, feeling drowsy and pain, with loadings ranging from 0.620 to 0.810. The emotional-sleep cluster comprised feeling nervous, irritability, sadness, difficulty sleeping and worrying.

Physical symptoms included shortness of breath, chills, palpitations and sweating. The self-image cluster contained skin changes, the sensation of not looking like oneself and hair loss; gastrointestinal symptoms comprised constipation, bloating and lack of appetite; and the neurological cluster consisted of dizziness and noise sensitivity. The analysis therefore grouped the three most prevalent complaints—fatigue, concentration difficulty and drowsiness—within one cluster, while showing that the wider burden extends into several partly distinct domains.

The team next constructed a 22-node Gaussian graphical model using the EBICglasso method in R. Nodes represented symptoms, while edges represented regularized partial correlations after accounting for the other symptoms. Thicker edges indicated stronger conditional associations.

The strongest estimated connections were between constipation and lack of appetite, between not looking like oneself and hair loss, and between shortness of breath and palpitations. Expected influence summarized a symptom’s net connectivity with the network, while bridge expected influence measured its net connections across the six PCA-derived communities. Fatigue ranked first in expected influence at 1.541, ahead of worrying at 1.169 and lack of appetite at 1.072.

Fatigue was also the leading bridge symptom, with a bridge expected influence of 0.430, followed by pain at 0.363 and difficulty sleeping at 0.309. These centrality estimates were supported by several robustness checks. One thousand nonparametric bootstrap resamples were used to estimate 95% confidence intervals for edge weights, and case-dropping bootstrapping produced correlation stability coefficients of 0.593 for strength and expected influence, indicating good stability.

Bridge strength and bridge expected influence had coefficients of 0.285 and 0.360, respectively, which the investigators categorized as acceptable. Networks based on automatic correlations, unadjusted Pearson correlations and Pearson residuals after adjustment for age and sex were highly similar, with edge-weight correlations from 0.978 to 0.993. Expected-influence rankings remained strongly correlated across models, with Spearman correlations of 0.905 to 0.965, as did one-step bridge rankings, at 0.820 to 0.913.

Fatigue and pain retained the two highest bridge expected influence positions in all three networks, although some weak cross-community edges disappeared after adjustment. Node-redundancy screening raised concerns about eight of 231 symptom pairs, but the researchers retained all 22 symptoms after reviewing their content. In 13 models that separately removed a flagged node, expected-influence rankings correlated with the primary model at 0.914 to 0.997, and bridge expected influence rankings at 0.934 to 0.997.

Fatigue remained first in expected influence in every model. It ranked first in bridge expected influence in 12 models and third when difficulty sleeping was removed; removal of feeling drowsy did not alter fatigue’s first-place ranking on either measure. These analyses indicate that the broad results were not driven by a single pair of highly similar items, though they do not remove the uncertainty inherent in an exploratory network.

The findings offer a structured way to assess the coexistence of symptoms rather than treating fatigue, mood, sleep and pain as isolated problems. In the studied population, fatigue combined high prevalence with the strongest measured connections to other symptoms, while worrying, lack of appetite, pain and difficulty sleeping occupied other prominent network positions. The authors say the results may support simultaneous screening of these domains and individualized planning, but they do not show that targeting a statistically central symptom improves outcomes.

They specifically caution that network centrality can identify candidates for clinical assessment and prioritization but cannot replace etiological evaluation, differential diagnosis or individualized decisions. Longitudinal and interventional studies would be needed to determine whether symptom networks change over time and whether management based on these clusters improves patient outcomes. Several limitations restrict how broadly the results can be interpreted.

The participants were recruited by convenience sampling from one tertiary hospital’s Traditional Chinese Medicine outpatient clinic, so people with different access to care, functional status or treatment preferences may be underrepresented. The sample was also relatively concentrated in middle adulthood and highly educated. Diagnosis relied on the 1994 Fukuda criteria, which do not require post-exertional malaise, and the symptom assessment did not specifically characterize post-exertional malaise or orthostatic intolerance.

As a result, the findings may not generalize to populations diagnosed under the 2015 Institute of Medicine criteria. The adapted Memorial Symptom Assessment Scale had high internal consistency in this sample, with a Cronbach’s alpha of 0.949, but it has not been formally validated in people with chronic fatigue syndrome. Self-report bias, cross-sectional design and adjustment only for age and sex further constrain causal or treatment interpretations.

The study therefore provides an exploratory map of symptom relationships in a defined clinic population, not a validated model of chronic fatigue syndrome or evidence of a treatment effect.

Subject of Research: Symptom clusters and network connections in adults with chronic fatigue syndrome

Article Title: Symptom clusters and symptom network characteristics in chronic fatigue syndrome patients: a cross-sectional study

Article References: Zi, A., Zhang, Y., Sun, M., Liu, T., Xu, Y., & Xu, M. (2026). Symptom clusters and symptom network characteristics in chronic fatigue syndrome patients: a cross-sectional study. Frontiers in Medicine, 13, Article 1955021. https://doi.org/10.3389/fmed.2026.1955021

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1955021

Keywords: chronic fatigue syndrome, ME/CFS, symptom clusters, symptom networks, fatigue, cross-sectional study, pain, sleep disturbance, Symptom, clusters, network, characteristics