Quality of life and socioeconomic indicators associated with survival of myeloid leukemias in Canada

Abstract Understanding how patient‐reported quality of life (QoL) and socioeconomic status (SES) relate to survival of acute myeloid leukemia (AML) and myelodysplastic syndrome (MDS) may improve prognostic information sharing. This study explores associations among QoL, SES, and survival through administration of the Euro‐QoL 5‐Dimension, 3‐level and Functional Assessment of Cancer Therapy‐Leukemia and financial impact questionnaires to 138 adult participants with newly diagnosed AML or MDS in a longitudinal, pan‐Canadian study. Cox regression and lasso variable selection models were used to explore associations among QoL, SES, and established predictors of survival. Secondary outcomes were changes in QoL, performance of the QoL instruments, and lost income. We found that higher QoL and SES were positively associated with survival. The Lasso model selected the visual analog scale of the EQ‐5D‐3L as the most important predictor among all other variables (P = .03; 92% selection). Patients with AML report improved QoL after treatment, despite higher mean out‐of‐pocket expenditures compared with MDS (up to $599 CDN/month for AML vs $239 for MDS; P = .05), greater loss of productivity‐related income (reaching $1786/month for AML vs $709 for MDS; P < .05), and greater caregiver effects (65% vs 35% caregiver productivity losses for AML vs MDS; P < .05). Our results suggest that including patient‐reported QoL and socioeconomic indicators can improve the accuracy of survival models.

vs $709 for MDS; P < .05), and greater caregiver effects (65% vs 35% caregiver productivity losses for AML vs MDS; P < .05). Our results suggest that including patientreported QoL and socioeconomic indicators can improve the accuracy of survival models.

INTRODUCTION
Acute myeloid leukemia (AML) and myelodysplastic syndrome (MDS) are malignancies of myeloid lineage, affecting around 10 individuals per 100 000, every year. Incidence rates nearly triple for those 70 and older and survival outcomes are poorer with increased age [1,2]. Decisions to offer curative treatments with allogeneic stem cell transplants (alloSCT) are guided by risk factors that are a function of increased age and/or disease progression [3,4]. Most other treatment options are, generally, noncurative in intent. The National Comprehensive Cancer Network (NCCN) guidelines consider cytogenetic abnormalities, leukemogenic mutations, co-morbidities, and geriatric conditions with established prognostic value [5][6][7][8]. Relatively less is known about the contribution of other risk factors that are independent of age or disease progression, such as quality of life (QoL) and socioeconomic status (SES). If other risk factors independently contribute to survival outcomes, then the accuracy of prognostic risk models may be improved through standardized data collection and incorporation into real-world models.
There is a growing literature that suggests patient-reported outcomes, such as QoL, may improve the accuracy of risk models used to predict survival outcomes. Incorporating patient-reported fatigue, for example, into the International Prognostic Scoring System for MDS (IPSS) has improved survival prediction [9]. The prognostic value of patient-reported health status for AML is less clear. One study in Italy reports a positive association between patient-reported QoL scores and survival of elderly adults with AML [10], whereas in Canada, the same association has not been made [11,12].
Missing or invisible prognostic data-such as patient-reported QoL-may be a source of error in evaluating chances of successful treatment if it is impactful. There is reason to believe that QoL both impacts survival and is a measurable outcome for treatment. Longterm AML survivors have reported poorer QoL compared to agematched members of the population without exposure to the disease or treatment [13]. The ability to evaluate new treatments also requires QoL data to estimate cost-effectiveness [14]. Because economic models depend on data from patients, and these QoL data do not exist, the ability to evaluate new treatments would be improved with more knowledge in this area.
Socioeconomic indicators are another unexplored and potentially impactfulmissing data. Studies in the Swedish population, have shown that elderly patients with geographic access to intensive treatment for AML live longer than those without [15]. If risks such as geographic access are found to be impactful on survival, then policy to address equitable access can be developed. If inequitable access to treatment remains an issue, however, the resulting survival data that are generated will inevitably suffer from selection bias. Black people with AML in the United States, for example, have less access to curative treatment for AML, inadequate alloSCT donor availability, and poorer outcomes after treatment-therefore, they are less likely to be able to contribute long-term survival data to predictive models and their health outcomes remain underrepresented [16]. Access to follow-up care is further restricted by tiers of insurance coverage, thus outcomes commonly overreport results for patients who are privately insured [17].
Knowledge on the contribution that patient-reported outcomes and/or SES has on survival and the resulting datasets generated could lead to improved modeling and better communication between clinicians and to patients and their families. In this study, we aim to explore the potential of these data to improve the accuracy of survival model-

Cytogenetic risk factors
Cytogenetic risk for AML was assigned according to current NCCN guidelines [4], with the addition of results from molecular testing.
Specifically, all participating institutions had adopted routine testing for mutations of FLT3 and NPM1 for intermediate-risk AML, and ckit for low-risk AML. Molecular testing was only undertaken if individuals were candidates for induction/CR1 consolidation treatment, that is, without significant age or comorbidities to preclude treatment with alloSCT. If molecular testing results for FLT3 or NPM1 for intermediate risk or ckit mutations for core binding factor t(8;21) or inv [16] AML were unavailable, an unconfirmed risk group status was assigned and included in the regression analysis as a dummy variable. All MDS-related cytogenetic changes according to the WHO 2008 classification scheme were considered to be adverse-risk AML.

Questionnaires
Following REB approval at each of the six study centers, study coor- the EQ-5D-3L were scored according to preference weights specific for Canada, to generate Canadian preference-based index utility values [18]. Individual FACT-LEU scores were calculated using directions from the questionnaire provider [19]. The percentage of individual income devoted to healthcare expenditures was determined by dividing the sum of each participant's self-reported expenses (specifically related to treatment), by their monthly income. Both individual and household income data were collected and the self-reported income measures were adjusted to account for income sharing within a household. The method for determining travel expenses, loss of productivity, and caregiver impacts is provided in full detail in the Supporting Information.

Instrument validity and reliability
The validity of the EQ-5D-3L and FACT-LEU was assessed for all participants, at each time point, through intra-instrument item correlations and inter-instrument total score correlations. The reliability of the FACT-LEU was evaluated with the coefficient for Cronbach's alpha and the corrected item total correlation scores for both instruments were calculated, at each time point. Correlation with EQ-5D-3L index scores and the visual analogue scale (VAS) measured the relationship of the index score to an individual's own perception of QoL.

Endpoints and statistics
Overall survival was defined as the time from enrollment in the study to either death or 24 months of follow-up, whichever occurred first. Event We used the least absolute shrinkage and selection operator (Lasso) method to enhance the accuracy and interpretation of the results from the Cox model [24]. The Lasso method allows all variables at multiple time points to compete simultaneously to add information to the model and selects the most competitive variables for inclusion. Postselection inference methods were used to compute valid P-values for the Lasso model [25]. A bootstrapped sensitivity analyses was undertaken with 1000 resamples for all models and the Lasso results.

Multistate modeling of QoL and survival outcomes
Multistate modeling was applied to the cohort data to simultaneously analyze QoL and survival outcomes. The method involves defining a set of health states that the cohort may experience and calculating the probability of transitions between them [20]. We distinguished health states with observed or anticipated differences in mean QoL scores, mortality, relapse rates, and/or healthcare costs.
The QoL data and transition probabilities (ie, the probability of moving between health states over a defined period of time) were calculated for each of the health states identified. The transition process following the initial baseline diagnosis and QoL data were modeled instantaneously and all other transitions were modeled with Weibull distributions, annually using a semi-Markov process starting from the date of diagnosis or to either death, relapse, or follow-up. Any post-remission relapse/transformation was modeled from the date of relapse or transformation to death or follow-up.
For each health state identified, the mean EQ-5D-3L index and FACT-LEU scores were calculated. Any missing QoL data were accounted for by imputing missing values on health states with at least 60% of the complete data. Use of an imputed dataset is critical to analyses of patient-reported outcomes in this disease area to adequately represent missing date due to adverse health status. The reason for missing data was recorded on the case report form and analyzed for each participant and time point. If the reason was related to poor health status (ie, not random), the imputed data were generated from the mean of the lowest quintile of the complete dataset, for each time point

Personal financial impacts
Mann-Whitney rank sum tests were used to test mean differences in out-of-pocket expenditures and productivity income losses due to time off from paid work and QoL differences related to the health states modeled for this cohort. We used Chi-squared or Fishers exact tests to distinguish frequencies of catastrophic healthcare spending, adverse impacts on caregiver productivity, and characteristics of nonresponding participants.

RESULTS
The study enrolled 188 potential participants prior to confirmation  Table 1). The EQ-5D-3L VAS scores were the least correlated with age of all the of the QoL measures and therefore were selected for inclusion in multivariable Cox regression models (Table 2). Two models, using baseline (T0) and month three (T1) EQ-5D-3L VAS scores, showed that the inclusion of this variable, and an SES indicator (high-school education or higher), improved the accuracy of prognostic information in the model, after adjustment for risk factors with known prognostic value, such as disease type, baseline platelet counts, and age above 60. The accuracy of both baseline and month three models improved when the results were stratified for males and females; however, the results from the models did not change when the sex-stratified models were analyzed separately, with this sample size. When all variables were allowed to The multistate modeling defined 10 finite health states for this cohort ( Figure 3 and Table 3 in the Supporting Information. Questionnaire response rates from participants were highest for the first three study time points (greater than 65%) and fell as low as 47% for T4 due to inconsistency in study staffing, when the funding of the study was under review. Response rates returned to 65% after the study review period for T5. The majority of missing questionnaire data were missing at random due to discontinuity of staffing (see Supporting Information). A comparison of

F I G U R E 3 Health states definitions
the characteristics of participants with missing responses showed that missing data were from younger participants and from one study center with the most participants and discontinuous staffing (P < .05).
AML patients who had remissions longer than 6 months and those with MDS were more likely to respond with scores of perfect health, indicating a ceiling effect for the instrument and potential aggreability bias for people with more desirable outcomes.
When compared to MDS, the mean out-of-pocket expenses (ie, costs for travel to clinics, accommodations, and uninsured prescription drugs) were higher for patients with AML at month three ($559 vs $239; P = .05) and high medical costs may persist over longer terms of follow-up; month six mean out-of-pocket expenses were $334 versus $129, P = .06. Nearly two thirds (67%) of all study participants reported out-of-pocket medical expenses that totaled more than 5% of their monthly income. Patients with AML who were in remission also reported greater productivity income losses at both T1 and T2 time points compared to all other patients. At T1 (3 months), patients with AML who had productivity income loss paid an uninsured aver-

DISCUSSION
The results from this study suggest that data on patient-reported QoL and SES can improve the accuracy of risk models and that scores from the EQ-5D-3L's visual analogue scale are the most important predictor to include in post-remission survival models. Health state "4" corresponds to any AML that relapsed within 3 months or was refractory to induction chemotherapy; health state "10" accounts for any postremission relapse or leukemogenic transformation from MDS to AML.
instruments due to the severity of the disease and the wide range of treatments and outcomes implied over time [21][22][23]. Our results are consistent with results from elderly patients with AML that do not report differences in QoL between intensive and nonintensive treatments with the FACT-LEU [24]. A better understanding of the association between QoL and prognosis could improve concerns over transparency and poor communication flow between patients with AML and their attending clinicians [25]. The income and productivity losses we observe agree with the literature on adverse patient-reported outcomes for AML in other countries and studies suggesting the need for financial support for people receiving treatment for AML [26][27][28].
Our study is limited by missing responses arising mostly from discontinuity in study staffing. Despite this limitation, sufficient data were captured to show socioeconomic and financial disparity in this cohort.
This is a novel accomplishment for this disease area, highlighting potential for including patient-reported outcomes in future studies. The response rates were similar to those reported from other studies of these diseases with more frequent missing values for individuals with poorer survival outcomes such as AML patients who did not receive induction therapy, or whose therapy did not result in CR1. There were, however, lower response rates for income and race/ethnicity questions than other baseline demographic reply items. Our results are therefore conservative in estimating the extent of disparity for AML patients and overrepresent outcomes for healthy individuals of higher SES.
The potential for QoL and SES variables to improve the accuracy of survival models warrants further attention. QoL outcome measures should be routinely obtained in clinical trials and the question on how meaningful those outcomes are needs to be robustly explored. Collecting information on SES from patients in research studies may improve population outcomes with information on disparity or potential gaps in unemployment insurance coverage. In Canada, any out-of-pocket expenditure greater than 5% of a person's net income is considered catastrophic and requires policy to protect those affected [29] The information from this study therefore suggests that Canadians with AML face more financial disparity than patients with similar conditions that requires policy attention.

CONFLICT OF INTEREST
The authors declare no conflict of interest.

DEDICATION TO STEPHEN COUBAN
Dr Couban, the principal investigator for the TFRI prognostic risk study, died before this publication. He was a proponent in the advancement of our understanding of patient-relevant outcomes and was enthusiastic about sharing these results with the hematology community. His influence on the lives of patients and grace of leadership guided our collaboration.