A prospective study of fatigue trajectories among in‐centre haemodialysis patients

Objectives Fatigue is common and debilitating among dialysis patients. The aim of this study was to understand the longitudinal trajectory of fatigue and consider sociodemographic, clinical, and psychological factors that are related to variation in fatigue levels over time. Design A prospective study of fatigue with yearly assessments over 3 years among prevalent in‐centre haemodialysis (HD) patients. Methods Fatigue severity was measured using the Chalder Fatigue Questionnaire and fatigue‐related functional impairment using the Work and Social Adjustment Scale. The trajectories of fatigue outcomes were examined using piecewise growth models, using length of time on dialysis as time. Sociodemographic, clinical, and psychological predictors of fatigue were assessed using linear growth models, using follow‐up time. Results One hundred and seventy‐four prevalent HD patients completed baseline measures, 118 at 12 months, 84 at 24 months, and 66 at 36 months. Fatigue severity scores decreased by 0.15 each year. Fatigue‐related functional impairment increased by 1.17 each year. In adjusted linear growth models, non‐white ethnicity was a significant predictor of lower initial fatigue severity (B = −2.95, 95% CI −5.51 to −0.40) and a greater reduction in fatigue severity of 1.60 each year (95% CI 0.35–2.36). A one‐point increase in damage beliefs was associated with a 0.36 increase in fatigue‐related functional impairment each year (95% CI −0.61 to −0.01). Conclusion Damage beliefs predicted an increase in fatigue‐related functional impairment over time. However, the data strongly suggested that fatigue outcomes vary by length of time on dialysis. Statement of contribution What is already known on this subject? At least 1 in 2 haemodialysis (HD) patients are clinically fatigued. Growing evidence is available on the important role of psychological factors in fatigue across chronic conditions. The contribution of psychological factors, beyond distress, to fatigue in HD has not been examined to date. What does this study add? Ethnicity played a role in the initial level of fatigue severity and over time. Damage beliefs predicted an increase in fatigue‐related impairment over time. Data strongly suggested that fatigue outcomes vary by length of time on dialysis.


Statement of contribution
What is already known on this subject?
At least 1 in 2 haemodialysis (HD) patients are clinically fatigued. Growing evidence is available on the important role of psychological factors in fatigue across chronic conditions. The contribution of psychological factors, beyond distress, to fatigue in HD has not been examined to date.
What does this study add? Ethnicity played a role in the initial level of fatigue severity and over time.
Damage beliefs predicted an increase in fatigue-related impairment over time. Data strongly suggested that fatigue outcomes vary by length of time on dialysis.
In most developed countries of the world, including the United Kingdom and the United States, the most common renal replacement therapy (RRT) is haemodialysis (HD), particularly at the start and among patients older than 65 (Hole et al., 2018;MacNeill et al., 2018). Fatigue is among the most common symptoms experienced by dialysis patients, affecting at least half, which often interferes with daily life and functioning (Horigan, 2012). Fatigue is a complex and subjective symptom that has been previously described as a distressing and persistent feeling of physical, emotional, and/or cognitive tiredness not proportional to exertion and not relieved by rest (Cella et al., 1998;Dittner et al., 2004;Ream & Richardson, 1996). There is also evidence to suggest that fatigue may contribute directly and indirectly to adverse clinical outcomes in dialysis patients (Bossola et al., 2015;Jhamb et al., 2009Jhamb et al., , 2011Koyama et al., 2010;Picariello et al., 2018). The Standardized Outcomes in Nephrology for HD workshop identified fatigue to be a core outcome (Evangelidis et al., 2017); however, a poor understanding of the pathogenesis of fatigue in this patient population impedes effective fatigue management strategies. Current management of fatigue in the form of pharmacological treatments appears ineffective (Bossola et al., 2011). Exercise-based interventions, although typically effective (Cheema & Singh, 2005;Segura-Ort ı, 2010), have been criticized for being unsuitable for patients with multimorbidities, disabilities, and in poorer health (Kosmadakis et al., 2010). Therefore, a solely biomedical approach may be insufficient in this setting. An alternative model that has been postulated in the literature to explain fatigue across long-term conditions (LTCs) is the cognitive-behavioural model (Wessely et al., 1998). According to this model, primary disease factors, such as anaemia, inflammation, and dialysis may trigger fatigue, yet, how patients react cognitively, emotionally, and behaviourally to the initial symptom of fatigue may worsen, maintain and perpetuate fatigue over time (Figure 1; Donovan et al., 2007;van Kessel & Moss-Morris, 2006). For example, once fatigue develops as a consequence of underlying disease processes, interpretations of fatigue as uncontrollable and lasting, and unhelpful thinking styles in response to fatigue, such as catastrophizing, symptom focusing, or perceiving fatigue as a sign of bodily damage (damage beliefs), may lead to increased anxiety and low mood and subsequently unhelpful behavioural responses, such as excessive rest in an attempt to control fatigue and reduce the damage perceived it is doing to their body. Excessive rest may in turn lead to deconditioning, poor sleep and physiological arousal related to anxious mood, and lead to the perpetuation of fatigue (van Kessel & Moss-Morris, 2006; Van der Werf et al., 2003;Vercoulen et al., 1998;Wessely et al., 1998). The importance of these factors in maintaining and perpetuating fatigue has been extensively documented in other conditions, including multiple sclerosis (MS), rheumatoid arthritis (RA), and cancer (Ali et al., 2017;Als en et al., 2010;Grayson et al., 2013;Jopson & Moss-Morris, 2003;van Kessel & Moss-Morris, 2006;Zalai et al., 2015). This conceptualization of fatigue has led to psychological treatments associated with clinically Figure 1. Cognitive-behavioural model of fatigue in chronic illness. This figure illustrates how fatigue can be triggered by primary disease factors, such as anaemia or haemodialysis in this patient group, and it can then be perpetuated and maintained by a cycle of negative beliefs, depression and/or anxiety, and maladaptive behavioural patterns. Over time, this self-perpetuating cycle may also lead to physiological consequences, such as increased inflammation, disruption of the sleep-wake cycle, chronic activation of the central and autonomic nervous systems as well as the endocrine system in response to stress, and deconditioning. This can further maintain fatigue and lead to poorer clinical outcomes. significant improvements in fatigue and related outcomes in MS, RA and cancer (Cramp et al., 2013;Kangas et al., 2008;van den Akker et al., 2016).
Nonetheless, psychological factors remain largely unexplored in fatigue among kidney failure patients (Picariello et al., 2017b(Picariello et al., , 2017. According to a recent study that examined changes in fatigue over 1 year among HD patients, the mean level of fatigue severity was relatively stable over time, but variations were evident in fatigue status among patients (Bossola et al., 2017), possibly indicative of distinct fatigue trajectories. These variations in status were largely unexplained by sociodemographic and clinical factors, except for the role of comorbidities, but psychological factors were not examined (Bossola et al., 2017). The cross-sectional data here suggested that dialysis patients' beliefs, mood, and behaviours contribute to fatigue, in particular, distress, negative beliefs about fatigue, and unhelpful behaviours (all-or-nothing and avoidance behaviours) were significant predictors of greater fatigue severity, explaining an additional 36% of variance (Chilcot et al., 2016). Similarly, distress, damage beliefs, avoidance behaviours, and fatigue severity were significant predictors of greater fatigue-related functional impairment, explaining an additional 44% of variance (Chilcot et al., 2016). In the current paper, the longitudinal findings are presented.
The overarching aim of this study was to understand the longitudinal trajectory of fatigue in HD patients and consider factors that are related to variation in fatigue levels over time, with the following objectives: Primary objectives 1. To model the longitudinal trajectory of fatigue symptoms in dialysis patients over a 3year period.
2. To evaluate the contribution of demographic, clinical, cognitive, behavioural, and emotional factors to fatigue over a 3-year period.

3.
To determine the level and trajectory of functional impairment caused by fatigue over a 3-year period Given the paucity of longitudinal fatigue data, an inductive approach to the modelling of trajectories was adopted. We hypothesized that negative beliefs about fatigue, distress, unhelpful cognitive and behavioural responses to fatigue will predict greater fatigue severity and fatigue-related functional impairment over time, after controlling for sociodemographic and clinical factors.

Study design
This was a prospective study assessing the trajectories and predictors of fatigue severity and fatigue-related functional impairment over 36 months. Assessments were conducted at yearly intervals (baseline, 12, 24, and 36 months). Ethical approval was granted by East Midlands-Leicester NRES committee (REC Reference number 14/EM/0037). Patient consent was obtained.

Participants
Patients were recruited from a renal outpatient service in England. All the patients were receiving thrice-weekly HD, with sessions generally lasting between 3.5 and 5 hr and relying on exclusively high flux membranes (polysulphone) with bicarbonate used as a buffer. The study ran between April 2014 and August 2017. A total of 279 patients were approached for study participation, with 174 providing informed consent and completing the baseline measures (62.4%). Details of the study including patient recruitment have previously been reported (Chilcot et al., 2016), and further detail on patient attrition is available elsewhere (Picariello et al., 2018). In brief, at 12 months 118 patients (87% of patients still alive, on dialysis, and able to complete the questionnaires) completed questionnaires, at 24 months 84 patients (98% of patients still alive, on dialysis, and able to complete the questionnaires) completed questionnaires, and at 36 months 66 patients completed questionnaires (99% of patients still alive, on dialysis, and able to complete the questionnaires). Although, at 36 months, only 37.9% of the baseline sample was available, attrition was driven by mortality and transplantation, with prevalence rates in line with the renal registry data (Byrne et al., 2018).
The inclusion criteria were as follows: (1) adults with a confirmed kidney failure diagnosis, (2) aged 18 or older, (3) treated with conventional thrice-weekly in-centre HD, (4) with a dialysis longevity >90 days, (5) able to speak or write in English, and (6) able and willing to provide informed consent. Patients were excluded if they had significant visual or physical impairment preventing completion of the questionnaires, had any known cognitive impairments, and had serious mental health conditions as noted in their medical history. Patients were not approached if they were judged to be unsuitable by the nursing staff, such as approaching end of life care, repeatedly unwell during the recruitment period or in the process of moving to peritoneal dialysis.

Data collection
Participants completed the study measures during a stable dialysis session at least 20 min after the initiation of treatment. Further detail on the sociodemographic and clinical data collected at baseline and at each data collection time-point is available elsewhere (Chilcot et al., 2016). Comorbidity was evaluated using the Charlson Comorbidity Index (CCI) at baseline (Charlson et al., 1987).

Psychological measures
All psychological measures were administered at each follow-up. The primary outcome variable, fatigue severity was measured using the Chalder Fatigue Questionnaire (CFQ; Chalder et al., 1993). This questionnaire has not been developed specifically for HD; therefore, it captures general fatigue. This scale consists of 11 items and scores are assigned for each response, using continuous scoring from 0 to 3. Higher scores represent greater fatigue severity. A cut-off of greater than 18 defines a fatigue case (Chalder et al., 1993;White et al., 2011). The measure has good psychometric properties in HD patients (Picariello et al., 2016).
The secondary outcome variable, fatigue-related functional impairment, was measured using the Work and Social Adjustment Scale (WSAS; Mundt et al., 2002). The following psychological predictor variables were measured using validated measures: fatigue perceptions using the Brief Illness Perceptions Questionnaire (BIPQ; Broadbent et al., 2006), cognitive and behavioural responses to fatigue via the Cognitive and Behavioural Responses to Symptoms Questionnaire (CBSQ; Ryan et al., 2018), and distress using the Hospital Anxiety and Depression Scale (HADS; Zigmond & Snaith, 1983). These measures have been previously described in detail (Chilcot et al., 2016). These measures demonstrated acceptable to good psychometric properties (i.e., internal reliabilities >0.70) within the current study.

Statistical analysis
Questionnaire scores were computed using proration to account for missing items within scales (Graham, 2009). This is described in more detail elsewhere (Picariello et al., 2018). In most cases, single items of a scale were left unanswered with the proportion of item-level missing data per variable without proration fluctuating between 6.3 and 14.9% at baseline, 3.4 and 7.6% at 12 months, 0 and 2.4% at 24 months, and 0 and 1.5% at 36 months. Given the small volume of missing item data and no systematic patterns of missingness, it is unlikely that multiple imputation at the item level would result in different findings compared to proration, despite criticisms in relation to proration (Mazza et al., 2015).
Approximately 33.3 and 36.6% of observations were missing and not due to death or transplantation in the fatigue severity and fatigue-related functional impairment models, respectively. Based on descriptive statistics and tabulated patterns, data were unlikely to be missing completely at random, particularly in the comparison of participants who died versus those who were transplanted. A survival analysis from this study cohort indicated that fatigue severity and fatigue-related functional impairment were significantly associated with an increased risk of death and decreased likelihood of transplantation (Picariello et al., 2018). Nonetheless, Little's MCAR test at each time-point indicated that the pattern of missing values did not depend on the data values (baseline v 2 (df = 275, N = 174) = 274.59, p = .51; 12 months v 2 (df = 178, N = 118) = 182.60, p = .39; 24 months v 2 (df = 46, N = 84) = 49.60, p = .33; 36 months v 2 (df = 46, N = 66) = 47.79, p = .40). Therefore, multiple imputation analyses were conducted, according to Sterne et al. (2009) using 20 imputations and including variables in the model associated with missingness, such as total time in the study, and censoring indicators. Multiply imputed data were deleted for patients after they had died or had a transplant; therefore, data were only imputed for 23% of the total sample. This dataset was used for the linear growth models.
Univariate associations between sociodemographic, clinical, and psychological variables with fatigue severity and fatigue-related functional impairment were examined using bivariate correlations (Pearson or Spearman depending on normal distribution of the data) for continuous variables; ANOVA comparisons for normally distributed categorical variables or the nonparametric Kruskal-Wallis H test. Gender, age, haemoglobin, and dialysis vintage, referring to length of time on dialysis, were included in all models, irrespective of their univariate associations with fatigue outcomes. Although significant correlations were evident between all psychological predictor variables and fatigue outcomes, they were selected for inclusion in the models based on univariate linear growth models.
Fatigue over time was modelled using mixed-effects regression models, which when applied to longitudinal data are often referred to as linear growth models (Rabe-Hesketh & Skrondal, 2008;StataCorp Longitudinal, 2005). Fatigue severity and fatigue-related functional impairment were used as continuous outcome variables in the models. These models are a generalization of the standard linear regression model that allow the analysis of repeated measures through the inclusion of random effects, which are estimates of the between-person variability. In the analyses presented here, we allowed for both a random intercept, capturing variability in the initial level of fatigue, and a random slope for the time variable, capturing variability in the rate of change in fatigue. Here, time was coded as months from the baseline assessment. Non-linear functions of time were assessed; however, standard criteria (chi-square and fit indices) indicated a linear rate of change as providing the most parsimonious fit to the data. The association between hypothesized predictors with the initial level of fatigue and the rate of change in fatigue was assessed by including a predictor variable and a predictor by time interaction term in the model.
Between-person and within-person variability was considered as part of exploratory data analysis, including an exploration of any time-dependent effects of age and dialysis vintage based on the mixed evidence on the association of these constructs with fatigue in this patient population. Exploratory analysis indicated that both fatigue severity and fatigue-related functional impairment varied non-linearly by dialysis vintage. The identified pattern in fatigue severity also makes sense physiologically, as well as psychologically, as patients in pre-dialysis care have an accumulation of fluid and toxins in their bodies, which improves with initiation of dialysis, but over time, the physiological consequences of dialysis and treatment burden may contribute to the increase in fatigue. Therefore, an alternative parameterization of time to the one above was implemented where time was coded as months from dialysis initiation, rather than time from baseline assessment. A piecewise linear growth model was estimated. This model extends the model described above to allow for different phases of change by the inclusion of more than one random time slope (Bryk & Raudenbush, 1992;Van Der Leeden, 1998). Specifically, we included two linear time slopes with a change point at 24 months allowing for the estimation of the initial level of fatigue, the rate of change in fatigue between 0 and 24 months from dialysis initiation, and the rate of change after 24 months from dialysis initiation. Predictors of fatigue outcomes were examined using linear growth models described earlier because more than half of the sample was receiving dialysis for 24 months or longer at baseline (N = 101, 58%).
Descriptive statistics and exploratory statistics were conducted in SPSS version 24. Models looking at fatigue over time and predictors of fatigue outcomes were conducted in STATA version 15. Data were converted from wide to long format and set as longitudinal (xtset) (Torres-Reyna, 2007) to use the mixed command for multilevel mixed-effects linear regressions (Box 1; Rabe-Hesketh & Skrondal, 2008;StataCorp Longitudinal, 2005). To apply the piecewise model, dialysis vintage was recoded to fit linear models in each period (Bryk & Raudenbush, 1992), where at each follow-up for slope 1 dialysis vintage of over 24 months was capped at 24, whereas for slope 2, dialysis vintage of 24 months or shorter was recoded as 0. Significance was set at p ≤ .05, using the standard a cut-off.

Sample characteristics
Baseline sociodemographic and clinical characteristics have been previously described elsewhere (Chilcot et al., 2016) and are available in Table 1 . A bar chart summarizing the frequency distribution of dialysis vintage at baseline in this sample can be found in the Supplementary File S1. At baseline, mean fatigue was 17.34 (SD = 6.54) and 82 patients (47.1%) scored above the threshold for clinical fatigue, scoring 18 or above on the CFQ. Fatigue severity ranged from 16.49 (SD = 6.12 at 36 months) to 18.26 (SD = 6.27 at 24 months) across follow-up time-points. At baseline, the mean fatigue-related functional impairment score was 18.5 (SD = 13.00), reaching a maximum of 22.37 (SD = 11.44) at 24 months. Descriptive fatigue data over time are presented in Table 2. Just over half of the participants who were fatigued at baseline remained fatigued at 12 months and slightly less than a quarter of participants remained fatigued by 36 months. Around 9 to 14% of participants who were not fatigued at baseline became fatigued at follow-up. Table 2 displays the descriptive characteristics of the psychological variables at each timepoint, based on the observed data. According to linear growth models, the yearly change in fatigue severity was À0.15 (p = .46, 95% CI À0.56 to 0.25), while there was a significant increase in fatigue-related functional impairment of 1.17 each year (p = .002, 95% CI 0.42 to 1.92). Using the MI dataset, the yearly change in fatigue severity was À0.11 (p = .68, 95% CI À0.63 to 0.41) and in fatigue-related functional impairment was 1.40 (p = .010, 95% CI 0.33 to 2.46). Similarly, relying only on participants who completed each follow-up (N = 66), the yearly change in fatigue severity was À0.24 (p = .31, 95% CI À0.71 to 0.23) and in fatigue-related functional impairment was 0.91 (p = .037, 95% CI 0.06 to 1.75). Using the MI dataset, without participants who died or were transplanted over the course of the study; produced similar results, with a yearly change in fatigue severity of À0.13 (p = .68, 95% CI À0.75 to 0.49) and in fatigue-related functional impairment of 1.39 (p = .041, 95% CI 0.06 to 2.73). Therefore, the findings are generally congruent.  Notes. BMI = body mass index; ESA = erythropoietin-stimulating agents; IDWL = intradialytic weight loss (weight lost during a dialysis session); URR = urea reduction ratio (marker of dialysis adequacy); CRP = C-reactive protein (marker of inflammation).  Based on unadjusted linear growth models, there were no significant changes over time in any of the psychological predictor variables, except for distress ( Table 2). The yearly reduction in distress was À0.46 (p = .038, 95% CI À0.90 to À0.03).

Predictors of fatigue severity over time
The results for each step of the model are available in Table 3. In the final model, including all the predictor variables (N = 14), only ethnicity was a significant predictor of the initial level of fatigue severity (B = À2.95, p = .024, 95% CI À5.51 to À0.40) and change in fatigue severity (B = 1.36, p = .009, 95% CI 0.35 to 2.36). Non-white ethnicity was associated with a 2.95 lower initial fatigue severity and a reduction in fatigue severity of 1.60 points each year. These coefficients represent clinically significant changes in fatigue severity on the CFQ, particularly taking into consideration the additive impact of ethnicity over time (Nordin et al., 2016). In addition, a one-point increase in haemoglobin was predictive of a small, but not clinically meaningful (Nordin et al., 2016), increase in fatigue severity of .02 each year (B = À0.05, p = .050, 95% CI À0.10, À0.00008).
Engaging in exercise at baseline was a significant predictor of lower initial fatigue severity, until distress was added to the model. Similarly, exercise status at baseline was predictive of a reduction in fatigue severity of 1.98 (step 1) and 2.13 (step 2) each year, respectively, until distress was added to the model. Distress was a significant predictor of the initial level of fatigue severity, where a one-point increase in distress was associated with a 0.30 increase in initial fatigue severity. This association ceased to be significant after negative beliefs about fatigue were added to the model. Lastly, a one-point increase in comorbidities (CCI) was associated with a 0.59 higher initial level of fatigue severity, but this was only marginally significant (95% CI À0.003, 1.18; step 2).

Predictors of fatigue-related functional impairment over time
The results for each step of the model are available in Table 4. In the final model, including all the predictor variables (N = 15), no significant predictors of the initial level of fatiguerelated functional impairment could be identified. On the contrary, change in fatiguerelated functional impairment was significantly associated with damage beliefs, after controlling for sociodemographic, clinical, and other psychological variables (B = À0.31, p = .041, 95% CI À0.61 to À0.01). A one-point increase in damage beliefs was associated with a 0.36 increase in fatigue-related functional each year. This is unlikely to be a clinically significant change in fatigue-related functional impairment (Zahra et al., 2014); however, over 3 years this would amount to an increase of 1.08 in fatigue-related functional impairment, where following therapy the mean change in WSAS has been reported to be À5.07 (Zahra et al., 2014).
Higher BMI at baseline was a significant predictor of greater initial fatigue-related functional impairment (step 1), yet ceased to be significant when clinical factors were added to the model. Being employed and engaging in exercise at baseline, on the other hand, were significant predictors of lower initial fatigue-related functional impairment. However, employment status was only marginally significant after distress was added to the model, while exercise status remained a significant predictor of initial fatigue-related functional impairment until negative beliefs about fatigue were added to the model. Additionally, in steps 1 and 2, engaging in exercise at baseline was predictive of a reduction in fatigue-related functional impairment by 1.19 each year, until distress was added to the model. Notes. Variables were entered into the model based on their univariate association with fatigue outcomes, to account for multiple tests an adjusted p-value cut-off of p = .002 was also used at baseline (Bonferroni correction 0.05/31), but with reduced power over time given attrition, the magnitude of the association was considered (r ≤ 0.30) regardless of the p-value. CI = confidence interval.
Bold values indicate p < .05. Notes. Variables were entered into the model based on their univariate association with fatigue outcomes, to account for multiple tests an adjusted p-value cut-off of p = .002 was also used at baseline (Bonferroni correction 0.05/31), but with reduced power over time given attrition, the magnitude of the association was considered (r ≤ 0.30) regardless of the p-value. CI = confidence interval.

Fatigue over time in HD 79
Being deemed unfit for a transplant at baseline was also predictive of greater initial fatigue-related functional impairment in step 2, but this association was only marginally significant (95% CI À0.09 to 10.31). Lastly, a one-point increase in distress was associated with a 0.69 increase in initial fatigue-related functional impairment, but this association ceased to be significant after negative beliefs about fatigue were added to the model.
The predictors of fatigue severity and fatigue-related functional impairment were generally consistent when including only participants who self-identified as White,Black African,Black Caribbean,or Black Other (N = 156; please see Supplementary File S4).

Discussion
The aim of this paper was to assess the trajectories of fatigue outcomes over 3 years and to identify predictors of these among prevalent (those receiving dialysis for 3 months or longer) in-centre HD patients.
Similar to previous estimates of the prevalence of fatigue among dialysis patients (Almutary et al., 2013), we found that approximately one in two patients (47.1%) could be deemed clinically fatigued at baseline, and many of those fatigued at baseline, who were followed up, remained fatigued at each subsequent follow-up.
The findings of the piecewise and linear growth models complement each other and reveal important nuances regarding fatigue change, suggesting that fatigue severity does not follow one linear trend of reduction over time. Fatigue severity appears to decrease within the first 24 months of dialysis, increasing thereafter with dialysis vintage.
While strong associations were evident across follow-ups between all psychological predictors and fatigue outcomes, this was not reflected in the linear growth models, which is also in contrast to the cross-sectional findings of this study (Chilcot et al., 2016). Psychological factors did not appear to play a significant role in fatigue severity, after controlling for sociodemographic and clinical factors. A lower initial level of fatigue severity and a reduction in fatigue severity each year were predicted by non-White ethnicity here. Contrary to expectations, an increase in haemoglobin was significantly associated with a small increase in fatigue severity each year. It is important to note that assessing the effects of time-dependent haemoglobin using a single measurement per time-point is less precise, and a time-averaged haemoglobin estimate over 3 to 4 months would be more appropriate.
Similarly, none of the sociodemographic, clinical, nor psychological variables emerged as significant predictors of the initial level of fatigue-related functional impairment. Damage beliefs in response to fatigue were the only significant predictor of an increase in fatigue-related functional impairment each year. The stepped approach adopted with the linear models, revealed some other factors that may play a role in fatigue outcomes, including employment status, exercise status, and distress. However, these factors were attenuated by the addition of psychological predictors to the models, suggesting some collinearity between the variables.
The findings of this study make a valuable contribution to our understanding of fatigue outcomes over time among dialysis patients, exceeding follow-up periods of previous longitudinal studies in this setting (Bossola et al., 2017;Jhamb et al., 2009Jhamb et al., , 2011). An important novel insight from this study is that fatigue is not merely associated with dialysis vintage, as previously suggested by some studies (Jhamb et al., 2011), but rather changes over the course of dialysis vintage. In the longitudinal study of fatigue over 1 year among prevalent HD patients by Bossola et al. (2017), fatigue severity also appeared relatively stable and was generally unexplained by sociodemographic and clinical factors when using the arbitrary follow-up time.
Contrary to expectations and previous findings on the important role beliefs and behaviours play in illness and symptom experience, including fatigue, across different LTCs (Hagger et al., 2017), only damage beliefs in response to fatigue independently predicted greater fatigue-related functional impairment over time here. The role of damage beliefs in fatigue-related functional impairment was also evident in the crosssectional data (Chilcot et al., 2016). Damage beliefs may be a proxy of actual damage occurring to the body as perceived by the patient; however, in line with existing literature, the majority of clinical factors, such as Hb, albumin, dialysis adequacy (URR), did not contribute significantly to fatigue severity and fatigue-related functional impairment here and the association between damage beliefs and fatigue-related functional impairment was above and beyond the role of clinical factors. Believing that fatigue is a sign of damage appeared particularly salient to renal patients and has been previously reported in a qualitative study, where patients often struggled to dissociate fatigue from the experience of the illness and treatment (Picariello et al., 2017a(Picariello et al., , 2017. Higher levels of damage beliefs, alongside catastrophizing and symptom focusing, were also observed in a sample of fatigued RA patients, possibly maintaining and perpetuating fatigue for this group as well (Ali et al., 2017).
In line with previous findings, non-White patients experience lower fatigue and are less likely to experience an increase in fatigue over time in this patient population (Jhamb et al., 2009(Jhamb et al., , 2011Kutner & Devins, 1998;Kutner et al., 2000Kutner et al., , 2005, possibly as a result of cultural differences in perception of the illness and treatment (Unruh et al., 2004), variations in coping styles employed (Bhui et al., 2011;Prelow et al., 2000;Triandis et al., 1984), or due to stigma associated with disclosing distress and maladjustment (Cooper-Patrick et al., 1997;de Crane & Spielberger, 1981;Thompson et al., 2004). Education and health literacy are also likely to be important confounders here. Lower educational attainment and low health literacy may influence fatigue via a poorer understanding of the illness and non-adherence, while patients with greater educational attainment may be more aware of the gravity and consequences of the illness. This further accentuates the complex aetiology of fatigue. It is also essential to consider this finding alongside the epidemiology of chronic kidney disease (CKD). The median age of patients on dialysis is highly dependent on ethnicity, specifically non-White patients tend to be younger (Hole et al., 2018;MacNeill et al., 2018). Therefore, age is also a likely explanation here. It is also yet to be determined whether the self-report instruments used here are cross-culturally valid.
According to the findings here, the timing of intervention delivery may be of relevance for patients on dialysis, as 24 months from dialysis initiation marked a change in the trajectory of fatigue severity. There is accumulating evidence in support of psychological interventions, such as cognitive-behavioural therapy for the management of fatigue in physical LTCs (van den Akker et al., 2016) and promising preliminary evidence is available in kidney failure (Picariello et al., 2017). However, what modifiable cognitive and behavioural factors are associated with fatigue trajectories over the span of dialysis vintage is yet to be defined.
To our knowledge, this is the first study to assess the contribution of fatigue-specific cognitive and behavioural factors to fatigue among in-centre HD patients. An additional strength of the study is its longitudinal design, enabling us to evaluate the trajectories of fatigue severity and fatigue-related functional impairment over 3 years and identify factors that may predict changes in fatigue, in an area that is largely dominated by cross-sectional research.
However, the current study has several important limitations. Firstly, although the sample consisted of 174 patients at baseline, it is still limited in size, particularly given the longitudinal study design, restricting the complexity of the analysis that could be conducted. In the light of findings in other LTCs, it would be clinically valuable to explore whether different clusters of fatigue trajectories exist in this patient population, using latent class growth modelling (Jung & Wickrama, 2008;Nagin & Odgers, 2010). Although high levels of attrition are inevitable in this patient population, the findings need to be interpreted with caution.
While the longitudinal design played a critical role in determining that fatigue changes with dialysis vintage, as prevalent HD patients were recruited into the study, this considerably limited our ability to use the natural trajectory of fatigue based on dialysis vintage to identify predictors of fatigue. Future studies should assess predictors of fatigue outcomes longitudinally using a larger sample of incident dialysis patients, those new to dialysis. However, such a study may prove challenging to conduct in practice, as the number of patients starting HD at any one centre per year is limited (Hole et al., 2018). Fatigue is also common in CKD patients who do not require RRT (Abdel-Kader et al., 2009); therefore, dialysis initiation may not necessarily reflect symptom onset.
Despite the longitudinal study design, predictor and outcomes variables captured individuals' internal states simultaneously, consequently reciprocal causation remains a problem (Singer & Willett, 2003). To circumvent this, asymptomatic CKD patients could be recruited and followed up beyond dialysis initiation; however, unless in the context of a large prospective epidemiological study, this is unlikely to be feasible in practice.
Although collecting data on-dialysis provided consistency and minimized patient burden, there is evidence of significant fluctuations in fatigue as a consequence of dialysis (Brys et al., 2019); therefore, to capture the full complexity of fatigue in this patient population, combining average and momentary assessments would be essential and would enable the identification of fatigue predictors both at the group level and at the individual level. Within-person methodology can also help to circumvent the limited target population of incident dialysis patients.
Due to concerns about patient burden, the role of important factors like subjective sleep quality, social support, and representation of the illness was not assessed here. Additionally, the use of the total BIPQ score precluded us here from exploring the role of individual fatigue perceptions.
Additional limitations, including the possibility of non-response bias, generalizability of findings to non-English-speaking samples, and sporadic availability of CRP from clinical records, have been previously addressed elsewhere (Chilcot et al., 2016;Picariello et al., 2018).

Conclusion
In conclusion, changes in fatigue outcomes over time may be more complex than originally anticipated, varying by the length of time on dialysis, as the data suggested here. This provides some indication of when treatment for fatigue may be effectively delivered, yet the targets of treatment need to be explored further using a sample of incident dialysis patients.