Cancer Symptom Clusters: An Exploratory Analysis of Eight Statistical Techniques

• Published in: Journal of pain and symptom management 2014-12, Vol.48 (6), p.1254-1266
• Main Authors: Aktas, Aynur, MD, Walsh, Declan, MSc, FACP, FRCP(Edin), Hu, Bo, PhD
• Format: Article
• Language: English
• Subjects: Aged; Anesthesia & Perioperative Care; Biological and medical sciences; Cancer; Cluster Analysis; Cohort Studies; Data Interpretation, Statistical; Female; Humans; Male; Medical sciences; Middle Aged; Multiple tumors. Solid tumors. Tumors in childhood (general aspects); Neoplasms - epidemiology; Neoplasms - physiopathology; Pain Medicine; Pharmacology. Drug treatments; Prevalence; Sensitivity and Specificity; Severity of Illness Index; statistics; symptoms; Tumors; United States - epidemiology
• ISSN: 0885-3924; 1873-6513
• DOI: 10.1016/j.jpainsymman.2014.02.006

Abstract Context Statistical methods to identify symptom clusters (SC) have varied between studies. The optimal statistical method to identify SC is unknown. Objectives Our primary objective was to explore whether eight different statistical techniques applied to a single data set produced different SC. A secondary objective was to investigate whether SC identified by these techniques resembled those from our original study. Methods We reanalyzed a symptom data set of 1000 patients with advanced cancer. Eight separate cluster analyses were conducted on both prevalence and severity of 38 symptoms. Hierarchical cluster analysis identified clusters at r -values of 0.6, 0.5, and 0.4. For prevalence and severity, the Spearman correlation and Kendall tau-b correlation, respectively, measured the similarity (distance) between symptom pairs. Sensitivity analysis of the prevalence data was done with Cohen kappa coefficient as a similarity measure. The K -means clustering method validated clusters. Results Hierarchical cluster analysis identified similar cluster configurations from the 38 symptoms using an r -value of 0.6, 0.5, or 0.4. A cutoff point of 0.6 yielded seven clusters. Five of them were identical at all three r -values used: 1) fatigue/anorexia-cachexia: anorexia, dry mouth, early satiety, fatigue, lack of energy, taste changes, weakness, and weight loss (>10%); 2) gastrointestinal: belching, bloating, dyspepsia, and hiccough; 3) nausea/vomiting: nausea and vomiting; 4) aerodigestive: cough, dysphagia, dyspnea, hoarseness, and wheeze; 5) neurologic: confusion, hallucinations, and memory problems. Regardless of the threshold, there were always some symptoms (e.g., pain) that did not cluster with any others. Seven clusters were validated by K -means analysis. Conclusion Seven SC identified from both prevalence and severity data were consistently present irrespective of the statistical analysis used. There were only minor variations in the number of clusters and their symptom composition between analytical techniques. All seven clusters originally identified were confirmed. Four consistent SC were found in all analyses: aerodigestive, fatigue/anorexia-cachexia, nausea/vomiting, and upper GI. Our results support the clinical importance of the SC concept.