Background

Lung cancer is the most common cancer in the world and accounts for 18% of cancer-related deaths [1]. Incidence of lung cancer continues to grow, and North America has the highest rates of diagnosis [1]. Mortality rates for this cancer type are high, with 1- and 5-year survival of 42 and 16%, respectively [2]. Lung cancer is associated with higher symptom burden compared to other cancers, and treatment-related symptoms are associated with clinical outcomes, including disease-free and overall survival (OS) and treatment success [3, 4]. Symptoms, such as fatigue, chest pain, and persistent cough frequently exist before diagnosis and may impact health-related quality of life (HRQOL) [3, 4].

HRQOL is a multidimensional construct defined as an individual’s self-reported sense of well-being as it relates to health-related event or illness and includes several domains, such as physical, functional, social, and emotional well-being [5–7]. HRQOL has been shown to be significantly associated with OS and healthcare services use in non-cancer diseases, such as arthritis and HIV [8–10]. A systematic review of the prognostic value of HRQOL for OS in cancer clinical trials found strong associations across 36 of 39 studies evaluated [8]. Twelve of these studies were conducted in lung cancer [8]. The review found that HRQOL was more predictive of OS than a clinician-rated performance status, which is routinely assessed in clinic [8]. Studies specific to lung cancer have confirmed significant relationships between HRQOL measured after cancer diagnosis and OS [11–13]. In lung cancer, the most commonly used instrument to capture HRQOL was the EORTC QLQ-C30, which was associated with statistically significantly 11–12% increases in the hazards of death [14, 15]. Global HRQOL measures, which have a narrower range of scores, were associated with large increases in the hazard of death with significant hazard ratios of 1.62 and 1.76 [16]. However, in all the studies included in the systematic review, HRQOL was measured after cancer diagnosis. Post-diagnosis (and often post-treatment) assessments may be confounded by exposure to diagnosis-related stress and treatment-related morbidities, which have been shown to negatively impact HRQOL [17]. To our knowledge, no study has examined associations between pre-diagnosis HRQOL and OS among individuals with lung cancer.

To fill this knowledge gap, we used a unique, population-based data set: the National Cancer Institute’s (NCI) Surveillance Epidemiology and End Results (SEER) cancer registry linked with data from the Center for Medicare and Medicaid Service’s (CMS) Medicare Health Outcomes Survey (MHOS). SEER provides detailed clinical data, and the MHOS includes HRQOL data for Americans 65 years and older (and disabled individuals under 65) enrolled in the Medicare Advantage Program. Our study’s objective was to determine if HRQOL measured before diagnosis could be predictive of OS and how relationships varied by HRQOL domain.

Methods

Data

The SEER consortium collects information on newly diagnosed cancer cases within SEER geographic regions covering 26% of the US population [18]. The MHOS is a questionnaire administered annually to 1000–1200 randomly selected beneficiaries from each managed care organization in the Medicare Advantage Program [18]. A baseline survey is administered along with a follow-up survey 2 years later [18]. Our study included 14 MHOS cohorts from 1998 to 2013. As Medicare Advantage plans are not represented in all SEER regions, there is an over representation from California, Detroit, and Seattle [18]. We obtained Institutional Review Board permission from the University of North Carolina at Chapel Hill.

Participants

We identified 7421 individuals with small-cell lung cancer (SCLC) and non-small-cell lung cancer (NSCLC), whose first SEER-confirmed cancer diagnosis occurred after baseline or follow-up MHOS. We only included individuals, whose lung cancer was the first and only cancer in SEER. Among these, 1131 (15%) were excluded due to lack of diagnostic confirmation of lung cancer. Some individuals completed multiple HRQOL assessments prior to diagnosis, in which case we took their closest assessment before diagnosis. Our sample consisted of 6290 diagnostically confirmed adults with incident lung cancer aged 65 years and older who completed an HRQOL assessment before diagnosis.

Covariates

Self-reported MHOS demographic characteristics collected pre-diagnosis included marital status, highest level of education completed, smoking status, and pre-existing health conditions. Age at diagnosis, sex, and race was also included. We adjusted for whether or not the MHOS was completed by a proxy (e.g., spouse or caregiver), as this may indicate worsened health status as well as if the survey was administered on paper or by telephone. We also controlled for SEER-reported cancer stage at diagnosis (local, regional, or distant), whether the lung cancer was NSCLC or SCLC, and treatments received (surgery and radiation).

Measures

OS

SEER-MHOS obtains information about death from death certificates. OS was calculated as number of months between MHOS survey and date of death (from SEER). Fifteen participants were diagnosed with lung cancer from their death certificate or via autopsy.

SF-36

The SF-36 was included in the MHOS from 1998 to 2005 [19]. The SF-36 has eight subscales: Physical Functioning, Role Physical, Bodily Pain, General Health, Vitality, Mental Health, Role Emotional, and Social Functioning [19]. The instrument also includes two summary scores: the Physical Component Summary (PCS) and Mental Component Summary (MCS). The eight subscales and PCS and MCS are normed scores with mean of 50 and standard deviation (SD) of 10 in the US general population with higher scores indicating better HRQOL [19].

MHOS cohorts from 2006 to 2013 used the Veterans Rand-12 (VR-12) instrument, which includes 12 items reflecting the eight SF-36 subscales and PCS and MCS [20]. We used an NCI algorithm to create comparable subscale and summary scores to combine data for those who completed the SF-36 in 1998–2005 with those who completed the VR-12 in 2006–2013 [21]. Subscales, MCs and PCS were normed with mean of 50 and SD of 10 in the US general population with higher scores representing better HRQOL.

Although there is no widely accepted standard minimally important difference (MID) for the SF-36 among individuals with cancer, we defined that an MID is a difference in group scores considered clinically relevant for patients or providers [22]. A review of the literature found SF-36 MIDs ranged from 0.2 to 0.5 SDs [23–25]. For our study, we used Cohen’s medium effect size of 0.5 of the SD [26]. Therefore, for the ten SF-36 normed scores, we used an MID of 5 point. We reported adjusted hazard ratios (HRs) associated with MIDs of 5 points. To convert our estimates, we took beta-coefficients for 1-unit increases, multiplied by 5 (the MID for these measures) and exponentiated the value. This allows us to present and interpret hazard ratios for MID in the SF-36 measures, which may be more clinically relevant than 1-point increases. Previous studies assessing associations between HRQOL and OS categorized PCS and MCS as below or above the mean of 50 [13, 27, 28]. We also dichotomized PCS and MCS scores to compare HRs to what was previously published.

General HRQOL

Our study included two single-item global questions. The first item asked, “In general, would you say your health is: excellent, very good, good, fair, or poor?” The second was, “In general, compared to other people your age, would you say that your health is: excellent, very good, good, fair or poor?” Excellent was used as our reference category.

ADLs

We examined responses to Katz’s basic ADLs (e.g., eating, bathing, dressing, getting in or out of chairs, walking, and using the toilet) [29]. Each item was phrased, “Because of a health or physical problem, do you have any difficulty doing the following activities without special equipment or health from another person?” Response options included, “No, I do not have difficulty”, “Yes, I have difficulty”, and “I am unable to do this activity”. Each ADL was assessed in a separate model, adjusting for demographic, comorbid, clinical, and treatment characteristics. “No difficulty” was the reference category.

Statistical analysis

Unadjusted comparisons of demographic, comorbid, clinical, and treatment characteristics between individuals who died and those alive at the end of follow-up were done using t tests and Chi-square tests. Model covariates were consistent with the previous SEER-MHOS studies [17, 22, 30, and 31]. Cox-proportional hazards models were used to examine associations between pre-diagnosis HRQOL and OS. Characteristics in Table 1 were adjusted for in all models. Although age at diagnosis is presented categorically in Table 1, it was treated as a continuous variable in analyses. We also adjusted for time from baseline HRQOL assessment to diagnosis as a continuous variable in our models. Our sample included 5107 deaths. Given the standard of ten or more events per covariate in a Cox-Proportional Hazards model, our models with 19 covariates were appropriate [32]. We computed adjusted HRs and 95% confidence intervals (CI) for each estimate. There were 3999 individuals who completed their MHOS more than 2 years before diagnosis. As these assessments may be less representative of HRQOL near diagnosis, we performed stratified analyses by those who completed assessments less or more than 2 years before diagnosis. We conducted additional sensitivity analyses restricting models to those with an assessment within 1 year of diagnosis. As there are distinct clinical differences between individuals with SCLC and NSCLC, we performed stratified analyses between the two groups. We also performed additional sensitivity analyses removing the 15 individuals who were diagnosed by autopsy or death certificate, but our estimates remained unchanged. Finally, we performed a sensitivity analysis removing HRQOL assessments that were completed by proxy respondents (9% of the sample), but the statistical significance and magnitude of the hazard ratios did not change. As such, as included an indicator for whether or not the HRQOL assessment was completed by a proxy in all models, but did not exclude these assessments from our models. Analyses were performed in SAS Version 9.3 with 2-sided statistical tests and a significance level of 5%.

Table 1 Demographic and clinical characteristics for individuals with lung cancer
Full size table

Results

Participant characteristics

Patient characteristics are presented in Table 1. Given our large sample size, some small differences between those who died and were alive at the end of follow-up were statistically significant, but, overall, we felt that age, race, SEER geographic region, education, marital status, and comorbid conditions were similarly distributed. Smoking status varied, as 47% of individuals alive at the end of follow-up identified as “never smokers” compared to 28% (p < 0.0001) among individuals who died. Fifteen percent of individuals who died had SCLC compared to 6% among those alive (p < 0.0001). Over 50% of those who died presented with metastatic disease, and 14% had surgery (50% had surgery in the alive group). Median time from HRQOL assessment to diagnosis was 28 months (interquartile range of 12–65 months).

SF-36 HRQOL

Mean PCS and MCS scores were 39.0 (SD 11.9) and 51.1 (SD 10.7), respectively. Adjusted HRs and p values for 5-point increments of PCS, MCS, and eight SF-36 subscales are shown in Tables 2 and 3. We present results for the overall cohort, those with an HRQOL assessment more than 2 years before diagnosis and for those with an HRQOL assessment within 2 years and 1 year of diagnosis (Table 2). We also present results stratified by NSCLC and SCLC (Table 3). As higher scores on SF-36 measures indicate better HRQOL, HRs in Tables 2 and 3 are below 1.0. That is, increases in HRQOL were associated with decreases in the hazard of death.

Table 2 Adjusted Cox-proportional model hazard ratios for short-form 36 measures
Full size table
Table 3 Adjusted Cox-proportional model hazard ratios for short-form 36 measure stratified by non-small-cell and small-cell lung cancers
Full size table
Table 4 Adjusted Cox-proportional model hazard ratios for activities of daily living measures
Full size table

Adjusting for demographic, comorbidities, clinical and treatment characteristics, a 5-point increase in PCS or MCS was associated with a 7 and 4% decrease in the hazard of death, respectively, in the overall cohort. Among individuals with an HRQOL assessment within 2 years of diagnosis, a 5-point increase in PCS or MCS was associated with a 5% decrease in the hazard of death. A 5-point increase in PCS was significantly associated with a 6% decrease in the hazards of death among those whose HRQOL assessment occurred more than 2 years before diagnosis. Results were similar when we restricted analyses to those with an HRQOL assessment within 1-year before diagnosis (Table 2). Having a PCS or MCS score above 50 was associated with 16 and 24% decreased hazards of death, respectively, and these results were similar when we subset analyses to individuals with an HRQOL assessment within 2 years of diagnosis. Five-point increases in SF-36 subscales were associated with 3–4% decreased hazards for the overall cohort, 5% among those with an HRQOL assessment within 2 years, 4% among those with an HRQOL assessment within 1 year, and 2–6% in those with an assessment more than 2 years before diagnosis. Compared to NSCLC, individuals with SCLC had larger HRs for PCS, Physical Function, Mental Health, Social Function, General Health, Bodily Pain, and Vitality domains (Table 3).

Global HRQOL

For the single-item measure, “In general, would you say your health is: excellent, very good, good, fair, or poor?” reporting poor (versus excellent) health was associated with a 74% increase in hazards of death (p < 0.0001) for the overall cohort. Reporting fair or good health was associated with 47 and 30% increases, respectively (p < 0.01). Smaller effects were observed for the second single-item question comparing one’s health to others. Reporting poor (versus excellent) health was associated with a 40% increase in the hazard of death. Reporting fair or good health was associated with 28 and 17% increases, respectively (p < 0.001). As worse HRQOL was associated with increased hazard of death, both global HRQOL questions had HRs above 1.0. We saw similar HRs in analyses restricted to individuals with HRQOL assessments within 2 years and 1 year of diagnosis (results not shown). HRs were similar between individuals with NSCLC and SCLC.

ADLs

Self-reporting that one was unable to complete any ADL compared to having no difficulty was significantly associated with the largest increases in hazard of death. In the overall cohort, being unable to bathe or dress oneself was associated with adjusted HRs of 1.89 (95% CI 1.47–2.44) and 1.88 (95% CI 1.36–2.59), respectively (Table 4). Magnitudes of HRs increased once we restricted models to individuals with HRQOL assessments within 2 years of diagnosis, but conclusions remained consistent (Table 4). When we assessed individuals with assessments more than 2 years before diagnosis, HR magnitudes generally decreased and some became insignificant. Being unable to bathe oneself, walk or use the toilet remained highly predictive of OS with adjusted HRs of 1.86, 1.54 and 1.64, respectively (Table 4).

Discussion

Our findings indicate that HRQOL collected before lung cancer diagnosis is a significant predictor for OS. Associations between poor HRQOL and increased mortality were seen across ten SF-36 measures, two single-item HRQOL measures, and six ADLs. Associations remained consistent when we restricted analyses to individuals with an HRQOL assessment within 1 and 2 years of diagnosis. Magnitude of adjusted HRs varied by HRQOL measures and domains. Differences in HR magnitudes between SF-36 domains and ADLs are mostly due to the fact that the SF-36 ranges from 0 to 100, whereas ADLs have three levels. One-unit increases in ADLs were associated with larger increases in the hazard of death compared to one-unit increases in SF-36 domains. We addressed this by interpreting MIDs for PCS, MCS, and eight subscales of the SF-36 instrument. ADLs also capture the basic physical functioning; thus, limitations or incapacity to complete activities signal poor health and are indicative of poor survival. Within the SF-36, physical health domains were more strongly associated with mortality risk than mental health domains.

Psychosocial SF-36 domains were also significantly associated with increased risk of death. The magnitude of the HRs for mental health domains and OS became larger when we restricted the cohort to individuals with an HRQOL assessment within 1 or 2 years of diagnosis, whereas they became a bit smaller for the physical health domains. This observation supports the value of routine screening for psychosocial HRQOL in clinical practice, as individuals with worse Role Emotional, Social Functioning, Mental Health, and Vitality were at increased risk of mortality. Furthermore, this association was consistent between NSCLC and SCLC patients, suggesting that psychosocial HRQOL may not be as related to specific disease characteristics as physical HRQOL.

A strength of our study is that we compared the prognostic value of single- and multi-item HRQOL measures on OS. Using single-item measures, we assessed overall perspectives of HRQOL, and since these are quick to measure and easy to report, they are clinically useful [11]. However, global measures do not allow for insights on granular, domain-specific HRQOL decrements identified with multi-item instruments [11]. Including results from both types of measures allowed for a comprehensive understanding of associations among different HRQOL domains and OS.

Our study findings are consistent with the previous work in lung cancer that examined associations between post-diagnosis HRQOL and OS. Ediebah et al. assessed associations between clinically meaningful 10-point increases in the EORTC QLQ-C30 and OS. They found increases/decreases in adjusted HRs ranging from 7% for physical functioning, 8% for pain, and 9% for social functioning [4]. Maione et al. found single-item global HRs ranging from 1.62 to 1.76, which are similar to our results (1.74 and 1.40) [16]. Another lung cancer study did not find ADLs to be predictive of OS, but found that Lawton’s Instrumental ADLs were associated with OS [16, 33].

Limitations

Our study only included adults with lung cancer 65 years and older, which limits generalizability to younger ages and other cancers. Individuals were in managed care plans, which may not generalize to Medicare fee-for-service beneficiaries. Evidence on whether or not managed care beneficiaries have better or worse health compared to Medicare fee-for-service beneficiaries is mixed [17, 34–36]. There are also limitations in the reliability of SEER treatment variables, such as radiation and surgery [37]. As receipt of chemotherapy is not included in SEER, we did not adjust for this in analyses. In addition, comorbid conditions are self-reported and do not have clinical confirmation. Approximately 9% of the MHOS were completed by proxies, and although we adjust for this in our analyses, it is a limitation.

The SF-36 was used between 1998 and 2005 and the VR-12 in 2006–2013. Although we used an NCI algorithm to combine scores, we recognize inherent limitations in using different instruments. In addition, we used the most recent MHOS for HRQOL assessments, but 50% had assessments more than 2 years before diagnosis. As such, stratified analyses between individuals with HRQOL assessments more and less than 2 years before diagnosis were conducted. Given associations between HRQOL and OS remained consistent regardless of when HRQOL was assessed, we find our findings striking, as patients may be aware of poor health well before diagnosis.

Conclusion

Our results have important clinical implications, as we observed a consistent pattern between pre-diagnosis HRQOL and OS as has been previously established with post-diagnosis HRQOL and OS. The current study does not address if HRQOL assessment timing (before or after diagnosis) is better. However, our results raise important consideration if routine monitoring of self-reported health status may have positive downstream effects. Routine HRQOL data collection in clinical practice has been shown to be feasible and acceptable to both patients and clinicians [38, 39]. Evidence also indicates that HRQOL monitoring enhances patient-physician communication and improves quality of care [39–42]. HRQOL may be used to identify individuals who are at a greater risk of death and who might benefit from targeted supportive HRQOL services [11]. Our study evaluated the prognostic value of various types of HRQOL assessments from domain-specific measures, questions about activities of daily living, and global HRQOL measures.

Allowing clinicians to intervene earlier when individuals report limited functioning or increased symptom burden may enable opportunities to mitigate effects of the underlying disease and perhaps, with earlier detection of cancer, potentially improve OS. However, this will need to be determined in a future study. Poor HRQOL ratings can be used for clinical follow-up for underlying physical and psychological causes [4, 11]. Consistent with the previous work, we recommend that HRQOL be routinely monitored throughout an individual’s interaction with the health care system [39]. Future studies should consider assessing associations between HRQOL changes (before and after cancer diagnosis) and OS to determine whether or not the same pattern is observed.