Trajectories of physical well-being among adults with acute myeloid leukemia

Acute myeloid leukemia (AML) is the most common myeloid malignancy, with an estimated 20 050 new cases in the United States in 2022 and a 5-year relative survival rate of 30.5%.¹ Physical well-being (PWB) is a facet of health-related quality of life (HRQoL) closely associated with physical functioning such as lower extremity function and daily walking.^2,3 Although most adults who have higher-risk AML (per Sorror et al composite model) rank HRQoL as less important than cure in their priorities for treatment choice, nearly half indicate QoL is more important than length of life.⁴ In addition to its importance as a patient priority, HRQoL is strongly linked to prognosis including survival, especially among patients aged >60 years.^4-7 

AML treatment regimens have differing effects on PWB. These regimens can be broadly categorized as intensive chemotherapy, typically administered in an inpatient setting, and lower-intensity chemotherapy, often given on an outpatient basis. Several studies have shown that although large proportions (>50%) of adults with newly diagnosed AML report PWB below normative values before starting intensive chemotherapy, most recover over several years of follow-up.^8,9 Although recent studies show overall HRQoL does not differ longitudinally by treatment type (intensive vs lower intensity) over 2 years of follow-up,⁴ historically, patients who received lower-intensity treatments have demonstrated better PWB.¹⁰ 

Although some studies have reported mean values and within-subject change for HRQoL among adults with AML, the literature lacks analyses exploring trajectories of PWB over time. Furthermore, few studies of PWB have (1) included both patients receiving intensive and those receiving lower-intensity chemotherapy and (2) adequately accounted for mortality-related loss to follow-up. Understanding how PWB changes over time in a sample with these characteristics will better reflect the experiences of patients in real-world practice settings and may help clinicians identify patients at high risk of adverse trajectories of PWB. In this study, we aimed to identify the trajectories of self-reported PWB among patients with AML and evaluate whether demographic and clinical characteristics (including receipt of intensive vs lower-intensity chemotherapy) are associated with these trajectories.

We conducted a secondary analysis of combined samples from 1 nonintervention (study 1) and 3 US supportive care clinical trials conducted between 2015 and 2019 (ClinicalTrials.gov identifier for study 2: NCT02975869¹¹; study 3: NCT03310918; and study 4: NCT03372291; each study has been approved by their respective institutional review boards). A new diagnosis of AML and the ability to provide informed consent were part of inclusion criteria for all studies. All patients in study 2 and study 4 were planned for or undergoing intensive chemotherapy treatment at baseline, whereas study 1 and study 3 also included patients on lower-intensity treatments. All patients were within the first year of treatment. Detailed inclusion criteria and description of interventions for each trial are provided in study characteristics of Table 1. We chose to combine patients across studies with similar patient populations and those that collected our outcome measure of interest to address limitations of previous research such as insufficient power for longitudinal analyses due to mortality-related loss to follow-up and poor generalizability due to single-site design.¹² 

Patient age, race/ethnicity, marital status, educational attainment, and income were obtained using standardized questionnaires at baseline.

European LeukemiaNet AML risk categories (favorable, intermediate, and adverse risk) based on the 2010 guidelines were reported by treating oncologists or abstracted from the medical records.¹³ Research staff determined whether patients were receiving or planned for intensive vs lower-intensity chemotherapy regimens. Depression and anxiety were measured using the hospital anxiety and depression scale, a valid and reliable screening tool for adults. This scale includes 7 items evaluating anxiety and 7 assessing depression, with scores ranging from 0 to 21 points (<8 points cut score indicates noncases).¹⁴ 

We chose the Functional Assessment of Cancer Therapy Physical Well-Being (FACT-PWB) subscale of the Functional Assessment of Cancer Therapy-General (FACT-G) measure as our outcome. The FACT-G measure and its subscales are valid and reliable self-reported measures of HRQoL among adults with cancer.¹⁵ The FACT-PWB consists of 7 questions with a range from 0 to 28 points (higher = better), with 2 to 3 points representing a minimal clinically important difference.^16,17 

We limited our sample to patients with FACT-PWB measures at ≥2 follow-up time points, who also had nonmissing values for baseline demographic and clinical covariates of interest. We compared the distribution of baseline covariates in our analytic sample with those excluded from the analysis due to missing data using χ² or Fisher exact test.

We report descriptive statistics for covariates of interest for the entire sample at baseline and by group as determined by group-based trajectory model (GBTM; see Modeling approach). When reported, unadjusted comparisons of covariates by group are tested using the SAS macro function COMPPROP, which uses a Tukey-style multiple comparison of proportions.¹⁸ 

To identify distinct trajectories of change in FACT-PWB over time, we used GBTM using the SAS macro function PROC TRAJ.¹⁹ GBTM are a class of discrete mixture models that estimate clusters of longitudinal data series and offer a flexible approach for modeling outcome data collected at varying time points. Our model specification used a censored normal distribution and an initial cubic polynomial trajectory shape of the curves. We additionally leveraged a GBTM extension to address nonrandom mortality related loss to follow-up for the FACT-PWB outcome.²⁰ A priori, we chose patient age, gender, marital status (married vs unmarried), educational attainment (college graduate vs some college or less), European LeukemiaNet risk category (favorable, intermediate, and adverse), receipt of intensive vs lower-intensity chemotherapy, and baseline depression, anxiety, and FACT-PWB as covariates to evaluate for association with group membership.

We determined the best fitting model with the ideal number of groups based on several criteria: (1) Bayes factor (approximated as 2 × ΔBIC) improvement between the saturated model (with greater number of groups) and the null model (with fewer); (2) group size; (3) average posterior probability of group assignment; and (4) clinical interpretability.^19,21 This last criteria was determined based on whether identified groups represented patient trajectories recognizable to oncologists in clinical practice. To ensure robust estimation of the Bayes factor, we used bootstrap with 500 repetitions (resampling with replacement using the SAS macro function BOOT).²² 

We conducted a sensitivity analysis excluding patients who died before contributing all expected measures of FACT-PWB. We present descriptive statistics for this group and results from GBTM. This analysis eliminated the nested model to weight against nonrandom mortality-related loss to follow-up and repeated the steps described above to select the GBTM model with the best fit. Importantly, we did not restrict this model to the same number of groups selected in the main analysis because (1) the sample size analyzed was substantially smaller and (2) we do not expect the same groups to be extracted given the distinct (more robust) composition of the patients included in the sensitivity analysis.

For clarity, results relating to groups from the main analysis are designated with prefix “m” (ie, group m1), whereas for those from the sensitivity analysis, we use the prefix “s” (ie, group s1). All P values are from 2-sided hypothesis tests with α of .05. Analyses were completed using SAS software version 9.4 (SAS Institute, Cary, NC).

Pooling patients from the 4 trials yielded an initial sample size of 405 (study 1, n = 99; study 2, n = 160; study 3, n = 88; study 4, n = 58). After eliminating patients with <2 measures of FACT-PWB during follow-up (n = 49) and those with missing values for predictor variables of interest (n = 13), the final analytic sample consisted of 343 patients. We compared the analytic with the original sample and found they were largely similar, with the exception that a greater proportion of those excluded were from study 3 (excluded, 50.0%; analytic, 16.6%; P < .001), were younger (excluded mean age, 64.6 years; standard deviation [SD], 12.7; analytic mean age, 69.5; SD, 12.1; P = .005), receiving lower-intensity chemotherapy (excluded, 62.9%; analytic, 28.8%; P < .001), and died during the follow-up period (excluded, 67.8%; analytic, 58.0%; P = .04; see supplemental Table 1).

The 343 patients included in this analysis had a mean age of 69.5 (SD, 12.1) years and were primarily male (n = 212 [61.8%]), White (n = 310 [90.3%]), and had a college degree (n = 181 [52.7%]). Over half (51.8%, n=178) had intermediate-risk AML, and 199 patients (58%) died while on study. Baseline FACT-PWB score was 19.9 (SD, 5.8), and patients contributed an average of 113.7 (SD, 63.6) days of follow-up (range, 11-200). See Table 2 for full results.

Based on our selection criteria (Table 3), the model with 4 groups was selected as the best model. The 4 groups of change in FACT-PWB identified in the final model are shown in Figure 1, and descriptive statistics of group characteristics are presented in Table 4. The largest group (m3: slight improvement/stable; n = 153 [45.0%]) had the lowest proportion of females (n = 43 [28.1%]), the lowest on-study mortality (n = 83 [54.3%]), and the second lowest baseline FACT-PWB (mean, 18.9; SD, 4.3). Conversely, the smallest group (m1: steep decline with recovery; n = 8 [2.4%]) was composed mostly of female patients (n = 5 [62.5%]) and those who died during follow-up (n = 6 [75.0%]) but started with a higher FACT-PWB (mean, 24.0; SD, 3.3) and lower depression and anxiety scores at baseline. We also present group size and average FACT-PWB across time points in Table 5. All 343 patients had baseline FACT-PWB scores, but only 183 of the sample (53.4%) contributed measures of the outcome for the final possible time point. Of the 160 (46.6%) with missing data for the final time point, 61 (38.1%) were missing data completely at random simply because this was not part of planned data collection for the participating trial (Table 1), whereas the remaining 99 (61.9%) were missing data due to loss to follow-up related to death, which was accounted for in the model estimation. When evaluated across groups, a comparison of multiple proportions found the percent of patients missing data due to death did not vary significantly by group. In addition, we trialed excluding the 8 patients in m1 from our analytic sample and found the remaining groups (m2, m3, and m4) were largely unchanged (results not shown).

Table 6 displays results from a multinomial logistic regression evaluating the independent association of each baseline factor with group membership. Group m3 was selected as the reference, because it represented the most common trajectory and relative stability of FACT-PWB score from baseline. Higher baseline FACT-PWB score was associated with increased relative odds of membership in groups m1 (odds ratio [OR], 1.57; 95% confidence interval [CI], 1.11-2.22; P = .01) and m2 (OR, 1.43; 95% CI, 1.25-1.62; P < .01), whereas higher FACT-PWB at baseline was associated 41% lower odds of membership in group m4 (OR, 0.59; 95% CI, 0.43-0.80; P < .01), all relative to group m3. Other baseline factors did not show statistically significant associations with group membership.

We also conducted a sensitivity analysis using GBTM to identify distinct trajectories of the outcome in a sample excluding patients with missing FACT-PWB data due to death (n = 99 [28.9%]; of the original analytic sample). The model with the best fit for this data identified 3 groups as shown in Figure 2: group s1 (n = 90 [37.0%]) showed early decline followed by partial improvement; group s2 (n = 137 [55.5%]) demonstrated a steady trajectory of slight improvement; and group s3 (n = 17 [7.5%]) initially improved and then stabilized as follow-up continued. Comparing group membership between the main 4-group model with the sensitivity analysis 3-group model, we found the majority of group s1 (96.7%) corresponded to group m2, 83.2% of those in group s2 corresponded to group m3, and all patients in group s3 corresponded to group m4 (supplemental Table 2). Supplemental Table 3 shows descriptive demographic and clinical characteristics for groups s1, s2, and s3. Interestingly, after excluding those with missing outcome data due to death, similar proportions of patients in all 3 groups had adverse-risk AML (range, 35.0%-35.6%) and received intensive chemotherapy (range, 71.5%-76.5%).

In terms of descriptive differences, although group s1 showed an improving trajectory early in follow-up and the highest baseline FACT-PWB score (mean, 24; SD, 2.9), it also had the fewest patients with favorable risk AML (n = 5 [5.6%]) and the greatest proportion of patients who died during follow-up (after contributing all expected FACT-PWB data, n = 42 [46.7%]). Conversely, group s3 started with the lowest FACT-PWB (mean, 9; SD, 3.7) but had the smallest proportion of patients who died during follow-up (n = 4 [23.5%]). Patients in group s3 were primarily female (n = 13 [76.5%]) and had higher levels of depression and anxiety at baseline. FACT-PWB scores are shown for the sensitivity analysis sample and by group over time in supplemental Table 4. The association of baseline factors with group membership among patients with complete data was largely consistent with results from the main analysis sample. Results from a multinomial logistic regression found baseline FACT-PWB score was the only characteristic independently associated with membership in groups s1 (higher FACT-PWB) or s3 (lower FACT-PWB) relative to group s2 (supplemental Table 5).

This analysis of longitudinal change in PWB among adults with AML identified 4 trajectories of change in FACT-PWB score over up to 200 days of follow-up. The largest group of patients experienced relatively little change in their PWB compared with baseline (group m3, 45%), followed by those who experienced a slight decline (group 2, 40.8%) and a smaller proportion of those who showed improvement (group m4, 11.8%) and sharp decline (group m1, 2.4%). These groups differed descriptively; a higher proportion of those in both groups m1 and m4 were female and were receiving intensive chemotherapy than those in the more stable groups. However, a multinomial logistic regression did not find that treatment intensity was significantly associated with group membership. Rather, the only demographic or clinical characteristic associated with group membership was that patients in improving trajectories had lower baseline FACT-PWB scores (group m4: mean, 9.9; SD, 3.8), whereas those who declined began, on average, with better PWB scores (group m2: mean, 24.0; SD, 3.2) than patients with relatively stable trajectories (group m3). Although the baseline average PWB score (19.9) was only slightly lower than normative values for adults with all cancer types (∼21.0),^23,24 the FACT-PWB does exhibit a ceiling effect that may be more pronounced in groups with higher baseline PWB (eg, group m2, in which 58.2% of scores were ≥24).²⁵ Another way of interpreting these results would be that we did not find any other baseline characteristics to be independently associated with PWB trajectory after adjusting for baseline PWB.

Existing literature has shown overall HRQoL among patients with AML often declines shortly after treatment initiation, but most patients demonstrate recovery over time,^9,26 although this recovery may be less robust than that experienced by patients with solid tumors.²⁷ However, this trend of improvement of overall HRQoL is less clear for PWB. Studies of adults with AML being treated with intensive chemotherapy found that although overall HRQoL as measured by the European Organisation for the Research and Treatment of Cancer scale improved in the years after diagnosis, the physical functioning subscale did not.^8,9 In contrast, Sorror et al reported similar FACT-PWB scores at baseline and improvement of HRQoL over 2 years in a sample of patients receiving both intensive and less-intensive chemotherapy.⁴ Our analysis focused on FACT-PWB similarly did not find treatment intensity to be independently associated with longitudinal trajectory. This may be related to the heterogeneity of disease course for patients receiving palliative, less-intensive treatment regimens, in that those responding well should demonstrate stable or improving PWB, whereas patients with disease progression may enter a terminal decline during the first 6 months of treatment.

Distinct from prior work, our analysis adds to the literature by (1) exploring distinct trajectories of change in FACT-PWB, rather than averages for an entire sample over time and (2) incorporating data from a substantial proportion of patients (58% of our sample) who died while on study. Although the largest group (group m3, n = 153 [45.0%]) identified showed stable or slightly improving FACT-PWB, we also found a distinct group of patients who slowly declined over the follow-up period (group m2, n = 141 [40.8%]). This result is complemented by a sensitivity analysis showing the most patients (63.0%) without missing data related to death improved their PWB during follow-up (group s2 [n = 137 (55.5%)], steady slight improvement; group s3 [n = 17 (7.5%)], early improvement, then stable). Of the 99 patients missing FACT-PWB measures related to death, most (n = 55 [55.6%]) were from groups experiencing decline (m2, n = 39; m4, n = 16). Together these suggest that average improvements in PWB reported among longer-term survivors may obscure underlying heterogeneity in real-world experiences of patients with AML, which often include terminal decline.

This secondary analysis is, to the best of our knowledge, the first study to explore risk factors associated with PWB trajectories among adults with AML. Factors associated with change in PWB have important potential clinical applications in identifying patients with high need who may be at risk for future decline. In studies of patients with solid tumors during their last year of life, GBTM have shown that factors such as lower levels of education and higher health care utilization are associated with a declining trajectory in overall HRQoL compared with maintaining a high HRQoL until death.²⁸ However, our analysis found that only baseline PWB was independently associated with trajectory of change. It is likely that predictors of PWB differ from those associated with overall HRQoL, and it is possible that demographic and clinical characteristics beyond those collected in our study are associated with longitudinal change in PWB. Geriatric assessment (GA), a combination of validated measures that assess specific domains (eg, function and nutrition) associated with adverse cancer–related outcomes, offers rich source of factors that may be associated with PWB. Although GA was originally developed for use among older adults with cancer, it is increasingly recognized as having applications among adults of all ages with cancer.²⁹ Measures of physical function often used in GA such as gait speed and activities of daily living may represent strong candidates for future studies exploring other potential predictors of PWB trajectory.

Our study has several limitations. We chose to pool data from studies with similar patient populations to address limitations of previous research including insufficient power and poor generalizability due to single-site design.¹² Although our sample was composed of patients with newly diagnosed AML, heterogeneity in interventions and engagement with study staff may have influenced FACT-PWB scores, decreasing the generalizability of these findings to patients who are not participating in supportive care studies. We chose to focus on trajectories of change in FACT-PWB rather than trajectories of total FACT-PWB scores, because it may be more clinically relevant to identify which patients are likely to decline regardless of their starting score. However, our findings of baseline PWB as the only significant predictor of trajectory group may be influenced by ceiling or floor effects of the FACT-PWB scale (given that patients with low scores were more likely to belong to a group with improving trajectory and patients with high scores were more likely to be in a group with decline). In addition, we did not have access to validated measures of physical functioning such as activities of daily living or physical performance tests such as gait speed. We were similarly limited in our characterization of socioeconomic status (education and marital status), which has been associated with HRQoL among adults with AML >1 year after diagnosis.³⁰ Finally, we lacked granular data on how treatment regimens changed over time and which patients progressed to hematopoietic stem cell transplant, which also may be predictive of longitudinal change in PWB.

Patients excluded from our analytic sample due to missing FACT-PWB measures were more likely to be receiving less-intensive treatments and die during the study. Although we attempted to account for mortality-related loss to follow-up in our models, our results may not represent the trajectories of patients receiving less-intensive chemotherapies. Future research is needed to confirm these trajectory groups, evaluate a broader set of measures for their association with group membership, and explore whether such trajectories are associated with other clinical outcomes.

HRQoL is an important outcome in the treatment of AML due to both its patient-centered nature and its association with other end points such as survival. Although prior studies have suggested HRQoL improves over time on average, we found this focus on the average HRQoL may obscure largely flat or declining trajectories that are revealed when investigating the subconstruct of PWB using modeling techniques that account for patients who died during follow-up. Future research including valid and reliable measures of physical function, such as those used in GA,³¹ is needed to identify baseline characteristics that may help predict future declines in PWB and allow for clinicians to implement targeted interventions to improve this important patient-reported outcome in this vulnerable population.

The authors thank Susan Rosenthal for editorial assistance.

Contribution: M.J.-B. analyzed and interpreted the data, wrote the manuscript, and approved the final article; M.B.S., W.C., and K.P.L. planned statistical analysis and interpreted the data, wrote the manuscript, and approved the final article; Y.W., Z.Z., M.L., J.D., K.B., and S.M. interpreted the data, wrote the manuscript, and approved the final article; and A.R.E.-J. collected and interpreted the data, wrote the manuscript, and approved the final article.

Conflict-of-interest disclosure: A.R.E-J. received support as a scholar in clinical research for the lymphoma and leukemia society; and has served as a consultant for GlaxoSmithKline, Incyte, AIM Pathway, as well as Tuesday Health. K.P.L. has served as a consultant to Pfizer and Seagen and has received honoraria from Pfizer. The remaining authors declare no competing financial interests.

Correspondence: Kah Poh Loh, Division of Hematology/Oncology, Department of Medicine, University of Rochester Medical Center, James P. Wilmot Cancer Institute, 601 Elmwood Ave, Box 704, Rochester, NY 14642; email: kahpoh_loh@urmc.rochester.edu.