The Value of a Symptom Cluster of Fatigue, Dyspnea, and Cough in Predicting Clinical Outcomes in Lung Cancer Survivors

Abstract

Context

Aggregates of concurrent symptoms, known as symptom clusters (SxCl), are reported to have prognostic capabilities beyond that of single symptoms alone. A SxCl of fatigue, dyspnea and cough has been delineated in a number of lung cancer cohorts.

Objectives

The objective of this study was to characterize this SxCl’s predictive value for important clinical outcomes relative to that of its component symptoms.

Methods

Analysis of an eight-year prospective cohort study that annually assessed 2405 patients with lung cancer for self-reported symptom burden, employment status, and physical activity with the Baecke questionnaire; symptom burden and overall quality of life (QoL) analysis was undertaken using nested Cox and generalized linear multilevel mixed models. Models were adjusted for longitudinally assessed demographics, cancer progression and tobacco use, and cancer progression.

Results

The SxCl, as well as its individual symptoms and symptom pairs, were all negatively associated with survival in Cox models of years 1–3 following diagnosis. Only in year 3 did the SxCl prognosticate survival (and then marginally) better than single symptoms or symptom pairs; fatigue was strongly associated (P<0.0005) with survival at all time points. The SxCl was not predictive of participants’ employment status, physical activity or QoL, whereas the presence of fatigue, dyspnea, or their combination was strongly associated with these outcomes.

Conclusions

Fatigue and dyspnea are strongly associated with poor clinical outcomes in lung cancer survivors; however, a SxCl that includes fatigue, dyspnea and cough as part as its components does not appear to significantly improve predictive capability.

Introduction

The idea that clusters of symptoms have prognostic capabilities beyond that of a single symptom alone is well accepted but may have been formally developed for the first time by Dodd, Miaskowski, and Paul¹ in 2001. The concept of symptom clusters (SxCls) has undergone continued development and has been recently examined by Kirkova and colleagues,² who draw a distinction between clinically and statistically defined clusters. This distinction, which has been noted by others as well,³ has face validity in that the symptoms within clinically defined clusters often share a common etiology (e.g., the gait, urinary, and cognitive dysfunctional changes associated with normal pressure hydrocephalus), whereas the factors responsible for associations between symptoms within statistically defined clusters are often more obscure.^{4, 5} Thus, while the utility of clinically defined clusters is often obvious, the practical relevance of statistically identified clusters remains less clear.

At present, a lack of data precludes an adequate assessment of the extent to which statistically defined symptom clusters can enhance patient care in a manner that optimizes outcomes, reduces cost, or improves prognostication. This is not to say that statistically defined clusters have limited value. On the contrary, as both death and functional decline are associated with statistically defined symptom clusters at six months to one year post-assessment in patients with cancer,^{1, 6, 7} they may inform practice in important and heretofore non-feasible ways. However, the longer term implications of these statistically based clusters beyond the first year of their identification are unknown. A persistent association with important outcomes at later dates would lend support to their practical utility.

Members of our research team recently identified a statistically defined SxCl of fatigue, dyspnea, and cough that was robust across different statistical discovery methods among 2405 lung cancer (LC) survivors enrolled in an three-year, prospective, longitudinal study.⁸ This SxCl is particularly intriguing in that it has been previously identified by others as occurring in LC survivors on the basis of both qualitative and quantitative methods. ^{3, 9} It is noteworthy also that the respiratory nature of this SxCl suggests a common, intuitively reasonable pathophysiology capable of influencing patients’ illness trajectories and clinical outcomes.

This paper describes the results of the use of nested, longitudinal models to estimate associations of the SxCl as well as its component symptoms with LC survival, employment status, physical activity, and overall quality of life (QoL).

Methods

Subject Recruitment

The recruitment and annual assessment of this study cohort has been previously described. ^{10, 11} In brief, all patients evaluated at the Mayo Clinic Rochester with a pathologically confirmed diagnosis of LC since January 1, 1999 have been prospectively followed. Twenty-five hundred (2500) survivors among this cohort agreed to complete a mailed self-report study instrument within six months of their diagnosis and each year thereafter until death. However, 4% of participants did not provide analyzable data; therefore, 2405 were included in the analyses. The study was approved by the Mayo Clinic Institutional Review Board.

Symptom Assessment

The questionnaire components utilized to define the SxCl of fatigue, dyspnea, and cough have been previously described. ⁸ Eleven-point symptom numerical rating scales were configured from 0 (as bad as it can be) to 10 (as good as it can be), lower scores thus indicating more intense levels of the symptoms.

Patient-Reported Clinical Outcomes

Quality of Life

Linear Analogue Self-Assessment (LASA) consisted of single items that were rated on a scale from 0 (as bad as it can be) to 10 (as good as it can be), such that lower scores similarly indicated a worsened QoL. Overall QoL, along with symptoms utilized to define the SxCl, were assessed in this manner. Data from a number of trials strongly support the validity and reliability of single-item LASAs.^12–14

Physical Activity and Employment Status

The Baecke Physical Activity Questionnaire is a validated measure of habitual vocational and avocational physical activity that has been widely used in epidemiologic studies and consists of work, leisure-time, sports, and total activity indexes.^{15, 16} The items “Are you currently employed/working?” and “I currently engage in regular physical activity,” both of which have binary (yes/no) response options, were utilized in the analyses described below.

Vital Status

Vital status of cohort members was established on an annual basis by reviewing death certificates, the Mayo Clinic’s electronic medical notes and registration database, and next-of-kin reports, as well as through the Mayo Clinic Tumor Registry and Social Security Death Index website.

Information from Electronic Medical Records

Research nurses abstracted data regarding survivors’ demographics, as well as tumor histology, staging, anatomical location, comorbidities and treatment from the Mayo Clinic electronic medical record. If survivors received treatment outside of the Mayo Clinic, copies of their treatment records were obtained and reviewed. ^17–19 Recurrent LC and the occurrence of a second primary LC were indicated with a binary “recurrence“ variable.²⁰

Statistical Analysis

Analysis extended from the point of entry into the data base till December 2008 or the time of death. Data obtained six months and one year after study enrollment were combined to generate a single average score as appropriate. To maintain consistency across models, all dependent variables as well as all symptom and SxCl covariates were binary. The ordinal QoL LASA was dichotomized at 5 since extensive prior work has demonstrated that a score ≤ 5 is indicative of a clinically significant degradation for QoL and is prognostically related to an increased mortality rate.^{21, 22} Survivors were considered SxCl(+) if they rated their fatigue, shortness of breath (SOB), and cough at 5 or above. Each symptom variable was similarly dichotomized at 5 to establish the values of symptom alone and symptom pair indicator variables. Seven indicator variables were assessed: the SxCl, fatigue + SOB, fatigue + cough, SOB + cough, fatigue, SOB, and cough. Only one indicator variable was included in each model. All models were adjusted for patient demographics; LC stage, histology, and treatment; medical comorbidities; smoking status; prior tobacco exposure; and LC recurrence.

Multiple modeling techniques were then applied to ascertain whether or not the SxCl actually provided greater information about clinical outcomes than individual or pairs of symptoms. Nested Cox proportional hazard models adjusted for baseline demographics, LC characteristics, and medical comorbidities were employed to characterize the prognostic capability of membership in the SxCl, as well as the relative contribution of single and paired symptoms. Models were created separately at each time point in order to capture differences in patient composition and prognostic factors. In particular, there were significantly more low stage patients in the cohorts that were 4 and 5 years out from their LC diagnoses. To examine the robustness of the results, Cox models also were created for each of the symptoms in the SxCl – fatigue, dyspnea, or cough – that included four mutually exclusive indicator variables (symptom not < 5, symptom < 5 with no other symptoms <5, exactly 2 symptoms in cluster < 5, SxCl present) and were adjusted as previously described.

A supplementary analysis using nested generalized linear multi-level and mixed models was utilized to model predictors of employment, the performance of regular exercise, and overall QoL. ^{23, 24} Only survivors < 65 years of age were included in models of employment status. Analyses were performed using SAS version 9 (SAS Institute Inc., Cary, NC). An α ≤0.05 was considered statistically significant.

Results

Table 1 lists the demographic, cancer and symptom characteristics of all participants at years 1 through 5 following LC diagnosis. Additionally, Table 1 lists the proportion of all survivors who were employed, exercised regularly, and reported their QoL to be ≤ 5, indicating lower QoL.

Table 1.

Demographic, Cancer and Symptom Characteristics of the Surviving Study Cohort Each of the First Five Years Following Lung Cancer Diagnosis

	Year >=1	Year 2	Year 3	Year 4	Year 5
N	1828	1244	941	738	482
% Symptom Cluster (SxCl(+))	15%	13%	14%	15%	15%
% Fatigue	50%	43%	45%	48%	50%
% Dyspnea	40%	39%	38%	44%	48%
% Cough	23%	19%	19%	20%	21%
% Fatigue + Dyspnea	32%	30%	31%	34%	39%
% Fatigue + Cough	18%	15%	15%	16%	17%
% Dyspnea + Cough	17%	15%	16%	17%	19%
% Male	53%	50%	50%	51%	48%
Mean age (SD)	65.9 (10.79)	65.2 (10.78)	64.8 (10.81)	64.1 (10.69)	63.4 (10.27)
LC Stage
I NSCLC	40%	52%	56%	59%	58%
II NSCLC	8%	8%	10%	11%	12%
III NSCLC/Limited SCLC	29%	26%	24%	22%	24%
IV NSCLC/Extensive SCLC	23%	13%	10%	7%	6%
N Eligible to work	1199	954	771	648	470
% employed	25%	24%	22%	24%	24%
N Eligible to work: excluding retired	325	251	197	171	126
% employed: excluding retired	89%	88%	86%	88%	88%
% engaging in regular physical activity	51%	61%	61%	60%	62%
% with QoL <= 5*	24%	20%	22%	24%	19%

^*Lower scores indicate a worse QoL

Implications of the SxCl Relative to Single Symptoms and Two-Symptom Pairs for Survival

Figure 1 presents adjusted survival curves for SxCl(+) and SxCl(−) participants at year ≤1 following LC diagnosis. On each of panels A through F in Fig. 1, survival curves for participants with and without each of the isolated symptoms or symptom pairs overlay the SxCl(+/−) curves. The presence of the SxCl, each of its component symptoms, and symptom pairs were strongly associated with diminished survival. The survival curves, though similar, suggest that the SxCl may be more strongly associated with diminished survival than any isolated symptom or symptom pair, excepting dyspnea and cough, at year ≥1. Cox proportional hazard models support this observation but suggest that the differences are small between the predictive capacity of the SxCl relative to symptom pairs or individual symptoms. Akaike information criteria (AIC) statistics for the Cox models at each time point are listed in Table 2. Only at year 3 did inclusion of the binary SxCl (+/−) variable (AIC = 2399.1) slightly enhance model fit relative to the inclusion of indicator variables for the presence of single symptoms or symptom pairs (AIC ≥2401.0). Table 2 further illustrates the diminishing association of the SxCl with survival over time, as reflected in the reduced hazard ratios (HRs) and increased P-values. Single symptoms and symptom pairs were almost equally or more strongly associated with survival at all time points than the SxCl. At year 4, fatigue was the only significant predictor of survival (HR 2.02, P=0.0005), and at year 5, the HR associated with fatigue increased to 4.32 (P<0.0001).

Figure 1.

Kaplan Meier curves comparing survival among patients with single symptoms and symptom pairs with patients experiencing the symptom cluster.

Table 2.

Results of Nested Cox Proportional Hazard Models at Each Time Point Adjusted for Demographics, Lung Cancer Characteristics, Smoking Status, Recurrence and Medical Comorbidities

		SxCl (+)	Cough	Fatigue	Dyspnea	Cough + Fatigue	Cough + Dypnea	Fatigue + Dyspnea
Year 1 N = 1828	Hazard Ratio	1.56	1.44	1.58	1.49	1.52	1.60	1.45
	95% CI	1.29 – 1.88	1.23 – 1.69	1.37 – 1.82	1.29 – 1.72	1.28 – 1.81	1.34 – 1.92	1.25 – 1.68
	p value	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001	<0.0001
	AIC	10579.8	10609.2	10641.3	10651.5	10579.1	10589.7	10643.2
Year 2 N = 1244	Hazard Ratio	1.42	1.58	1.99	1.59	1.55	1.54	1.57
	95% CI	1.04 – 1.93	1.22 – 2.05	1.60 – 2.48	1.27 – 1.98	1.16 – 2.06	1.16 – 2.05	1.24 – 1.99
	p value	0.0279	0.0005	<0.0001	<0.0001	0.0031	0.003	0.0002
	AIC	4210.8	4204.1	4218.8	4240.6	4207.2	4207.2	4243.5
Year 3 N = 941	Hazard Ratio	2.07	1.82	2.16	1.67	1.96	1.89	1.93
	95% CI	1.44 – 2.99	1.30 – 2.56	1.62 – 2.88	1.26 – 2.22	1.36 – 2.82	1.33 – 2.68	1.43 – 2.59
	p value	<0.0001	0.0005	<0.0001	0.0004	0.0003	0.0004	<0.0001
	AIC	2399.1	2429.3	2422.7	2454.7	2401.0	2428.7	2421.1
Year 4 N = 738	Hazard Ratio	1.43	1.20	2.02	1.25	1.40	1.38	1.48
	95% CI	0.83 – 2.46	0.75 – 1.94	1.36 – 3.01	0.83 – 1.86	0.85 – 2.29	0.82 – 2.30	0.98 – 2.23
	p value	0.1968	0.4467	0.0005	0.2841	0.1891	0.2261	0.0607
	AIC	1234.3	1235.5	1248.4	1259.8	1234.3	1234.7	1257.4
Year 5 N = 482	Hazard Ratio	1.46	1.18	4.32	1.79	1.48	1.29	2.44
	95% CI	0.74 – 2.88	0.63 – 2.19	2.37 – 7.88	1.02 – 3.13	0.77 – 2.87	0.68 – 2.43	1.40 – 4.25
	p value	0.2753	0.6034	<0.0001	0.0411	0.2428	0.4410	0.0016
	AIC	589.5	612.5	586.5	608.3	600.3	601.2	591.6

SxCl = Symptom cluster, AIC = Akaike information criteria.

Modeling results did not differ significantly when four mutually exclusive indicator variables were included in the survival analyses from the results of the nested Cox models described above. The HRs and P-values associated with no symptom ≤ 5 (lower score indicates more intense symptom), fatigue, fatigue + either dyspnea or cough but not both, and the SxCl are listed in Table 3. The HRs were not significantly different for the symptom pairs or the SxCl relative to fatigue alone. Similar patterns were observed for both dyspnea and cough at all time points excepting one for each symptom. Relative to symptom alone, the SxCl had a significantly higher HR than dyspnea (HR 2.8, P = 0.02) at year 3 and cough at year 1 (HR 1.8, P = 0.01).

Table 3.

Sample Sizes, Hazard Ratios, and P-Values at Each Year Adjusting for Baseline Variables in Cox Models That Include Four Mutually Exclusive Indicator Variables

Model	Year <=1	Year 2	Year 3	Year 4	Year 5
No Fatigue Symptoms	N=916	N=706	N=519	N=383	N=241
	0.65	0.48	0.47	0.42	0.17
	<0.0001	<0.0001	0.0003	0.0014	<0.0001
Fatigue ≤5* and no other symptoms	N=255	N=143	N=118	N=93	N=48
	reference	reference	reference	reference	reference
	group	group	group	group	group
Fatigue ≤5* and either Cough ≤5* or SOB ≤5* but not both	N=391	N=235	N=171	N=154	N=119
	0.95	0.92	0.88	0.80	0.72
	0.6360	0.6791	0.5887	0.4485	0.4318
Symptom Cluster	N=266	N=160	N=133	N=108	N=74
	1.20	0.86	1.26	0.86	0.50
	0.1570	0.4641	0.3534	0.6490	0.1352

^*Lower score indicates more intense symptom

Implications of the SxCl for Employment Status, Performance of Regular Exercise and Overall QoL

Tables 4–6 list the AIC statistics for nested generalized linear multi-level and mixed models of employment status, engagement in “regular physical activity,” and overall QoL, respectively. Among patients who were ≤63 years old at the time of LC diagnosis and could be reasonably expected to continue or resume gainful employment up to five years after diagnosis, the presence of dyspnea (P=0.003) and fatigue (P=0.05) were associated with not working. The presence of both dyspnea and fatigue (P=0.006) also was associated with not working, but was less informative than dyspnea alone. Additionally, both dyspnea (P<0.0001) and fatigue (P<0.0001), as well as the symptom pairs fatigue + dyspnea (P<0.0001) and fatigue + cough (P=0.04) were negatively associated with engagement in “regular physical activity.” A similar pattern was noted for overall QoL in that fatigue (P=0.0004), dyspnea (P<0.0001), and these symptoms together (P<0.0001) were associated with a QoL rating ≤5. In this case, the symptom pair was marginally more informative per the AIC that the individual symptoms. The SxCl was not associated with any of these outcomes.

Table 4.

Results of Nested Generalized Linear Multi-Level and Mixed Models of Employment Status Adjusted for Demographics, Lung Cancer Characteristics, Smoking Status, Recurrence and Medical Comorbidities

	N	Odds Ratio (95% CI)	p-value	AIC
Symptom Cluster	754	0.78 (0.48 to 1.28)	0.329	955.73
Cough	754	0.97 (0.62 to 1.50)	0.873	956.79
Fatigue	754	0.70 (0.49 to 1.00)	0.050	952.27
Dyspnea	754	0.58 (0.41 to 0.84)	0.003	946.85
Fatigue and Dyspnea	754	0.60 (0.41 to 0.86)	0.006	947.79
Fatigue and Cough	754	0.79 (0.49 to 1.26)	0.316	955.67
Dyspnea and Cough	754	0.90 (0.56 to 1.45)	0.586	956.61

SxCl = Symptom cluster, AIC =Akaike information criteria.

Table 6.

Results of Nested Generalized Linear Multi-Level and Mixed Models of Overall Quality of Life Adjusted for Demographics, Lung Cancer Characteristics, Smoking Status, Recurrence and Medical Comorbidities

	N	Odds Ratio (95% CI)	p-value	AIC
SxCl	718	1.42 (0.72 to 2.77)	0.310	419.30
Cough	718	1.35 (0.71 to 2.55)	0.358	419.46
Fatigue	718	4.04 (1.87 to 8.70)	0.0004	402.78
Dyspnea	718	3.90 (2.01 to 7.55)	<0.0001	401.40
Fatigue and Dyspnea	718	3.93 (2.12 to 7.28)	<0.0001	399.51
Fatigue and Cough	718	1.32 (0.69 to 2.55)	0.403	419.61
Dyspnea and Cough	718	1.50 (0.78 to 2.88)	0.229	418.90

SxCl = Symptom cluster, AIC = Akaike information criteria.

Discussion

This work assessed the utility of a frequently identified SxCl in predicting clinical outcomes among lung cancer survivors and surprisingly found that its utilization offered limited to no additional information beyond that provided by isolated symptoms and symptom pairs. It is significant that our work benefited from access to a large, well-established database and involved a cluster of symptoms (fatigue, dyspnea and cough) that is not only intuitively reasonable but one that has been identified by others in LC survivors on the basis of qualitative and quantitative methods.^{3, 9}

As such, several findings seem to be of some interest. First, although the presence of the SxCl at baseline was strongly predictive of shorter survival, it was not significantly more so than symptom pairs. Second, the prognostic significance of the SxCl diminished over time whereas the presence of single symptoms, particularly fatigue, remained strongly predictive of survival. Third, single symptoms and symptom pairs were predictive of patients’ reports of employment status, physical activity, and overall QoL while the SxCl was not.

In essence, our results, while in accord with reports of poor prognoses in LC patients manifesting SxCls during the first years, ^{7, 25, 26} highlight what appears to be a more limited predictive capability than that of the single symptoms of fatigue and dyspnea in the following years. More specifically, while the SxCl offered marginally better prognostic capacity over single symptoms in the first year (HR 1.6, P<0.0001, AIC 10579.8), it was no better than the symptom pair of cough and fatigue (HR 1.5, P<0.0001, AIC 10579.1). Further, the SxCl’s predictive capacity, unlike that of single symptoms such as fatigue, dwindled to the point that it was not associated with survival in years 4 and 5. In contrast, fatigue remained a significant predictor of survival at all time points. The prognostic significance of fatigue has been previously described^27–29 but, as far as we know, not among survivors as far out as eight years following LC diagnosis.

The superior prognostic capacities of single symptoms and the symptom pair of dyspnea and fatigue, relative to the SxCl’s lack of predictive capability, were particularly evident in their associations with employment status, physical activity, and overall QoL. The single symptoms of dyspnea and fatigue were both strongly associated with poor outcomes. Potential mechanisms for this finding must remain speculative, but may lie in their being markers for more aggressive past cancer treatment or cumulative medical comorbidity.

Limitations

Our criterion of a symptom intensity rating of at least 5/10 for inclusion in the SxCl, though empirically based on extensive prior work, ^{23, 24} was nonetheless stringent and our findings cannot be generalized to clustering of more mildly experienced symptoms in LC or other cancer populations. Though some members of the cohort were receiving anti-cancer therapies at the time of follow up, many were not. Therefore, SxCls related to active anti-cancer treatment, which may have greater prognostic significance for clinical outcomes, are not well represented in this study’s cohort.

Of note, controversy persists as to how many co-occurring symptoms constitute a cluster.³ Both two³⁰ and three³¹ symptoms having been proposed as threshold criteria. Our SxCl discovery methods⁸ revealed a strong association between all three symptoms; however, associations between fatigue and dyspnea were consistently stronger. Even if we were to consider the symptom pair of dyspnea and fatigue as a cluster, our conclusions would not differ significantly. Single symptoms, particularly fatigue, remain the same as or only slightly less predictive than the pair.

Two of the four outcomes were binary and objective, i.e., vital and employment status, while two were more subjective in nature, i.e., participation in regular physical activity and overall QoL, and the fact that we assessed these latter outcomes with single items may raise concern. With respect to QoL, extensive validation of 11-point numerical rating scales has been described among cancer populations. ^{12, 13} We acknowledge that use of a single item to assess physical performance has more limited precedent; however, psychometric evaluation of this approach suggests acceptable reliability and concurrent validity,^{32, 33} and prior research suggests the single items are reasonable proxy estimates of the degree of physical activity in a population.³⁴ We would have preferred to utilize the Baecke questionnaire summary scores, but the amount of non-random missing data would have potentially biased our results. Reassuringly, the prognostic performance of the SxCl, symptom pairs, and isolated symptoms was very similar for all four assessed outcomes, subjective and objective alike.

Conclusions

The presence of a SxCl comprising dyspnea, fatigue, and cough appears to provide little or no improvement in prognostic capacity relative to single symptoms and symptom pairs. This appears particularly true beyond the first year after LC diagnosis. Although further work is necessary, fatigue or dyspnea, alone or together, may be sufficient to determine important outcomes and warrant attention in cancer survivors.

Table 5.

Results of Nested Generalized Linear Multi-Level and Mixed Models of Performance of Regular Physical Activity Adjusted for Demographics, Lung Cancer Characteristics, Smoking Status, Recurrence and Medical Comorbidities

	N	Odds Ratio (95% CI)	p-value	AIC
Symptom Cluster	694	0.68 (0.42 to 1.10)	0.114	946.00
Cough	694	0.79 (0.52 to 1.21)	0.284	947.42
Fatigue	694	0.49 (0.35 to 0.68)	<0.0001	929.72
Dyspnea	694	0.43 (0.31 to 0.61)	<0.0001	924.90
Fatigue and Dyspnea	694	0.46 (0.32 to 0.66)	<0.0001	929.43
Fatigue and Cough	694	0.61 (0.38 to 0.96)	0.040	943.90
Dyspnea and Cough	694	0.77 (0.48 to 1.22)	0.260	947.29

SxCl = Symptom cluster, AIC = Akaike information criteria.