Introduction

Colorectal cancer (CRC) is a major health concern, representing the third most commonly diagnosed cancer worldwide and the second leading cause of cancer mortality in the developed world [1, 2]. Following significant improvements in the treatment of CRC as a result of advances in systemic chemotherapy [3], use of agents targeting the vascular endothelial growth factor and epidermal growth factor receptor (EGFR) now also have an established role in patients with metastatic disease [46].

Panitumumab is a fully human monoclonal antibody directed against the EGFR. Its clinical efficacy in metastatic CRC (mCRC) was confirmed in an open-label, phase 3 trial, in which patients receiving panitumumab plus best supportive care (BSC) had significantly improved progression-free survival (PFS) compared with those receiving BSC alone (hazard ratio [HR] 0.54; 95% confidence interval [CI], 0.44 to 0.66) [6] in patients with chemotherapy-refractory mCRC. Mutations in KRAS—a gene encoding a GTPase that mediates downstream intracellular EGFR signaling—are found in approximately 35–45% of CRC tumors [7] and have been associated with a lack of response to EGFR inhibitors when used for mCRC [8]. Evaluation of the predictive role of KRAS mutational status in the response to panitumumab in this phase 3 monotherapy study identified that its beneficial effect was restricted to patients whose tumors had wild-type KRAS. In patients with wild-type KRAS mCRC, panitumumab plus BSC was associated with a statistically significant improvement in median PFS compared with BSC alone (HR, 0.45; 95% CI, 0.34–0.59 [9]). In mCRC, the clinical use of panitumumab (like cetuximab) is therefore now restricted to patients with wild-type KRAS tumors. Despite recent advances in the treatment of mCRC provided by targeted agents, most patients will eventually die of the disease. The goal of therapy in this setting is, therefore, to delay disease progression, control symptoms, and maintain health-related quality of life (HRQoL) for as long as possible [10, 11]. As such, patient-reported outcomes are particularly important to patients with metastatic disease, and the potential for benefit related to these outcomes is a key consideration in the choice of therapy [12].

During the phase 3 trial of panitumumab monotherapy, HRQoL and CRC symptom data were collected. Patient-reported outcomes from the trial were briefly reported in the primary efficacy paper but only for the overall patient population and not by KRAS tumor status [6] (online appendix only). Since panitumumab is only indicated for patients with wild-type KRAS tumor status, the available patient-reported outcomes data do not, therefore, reflect the findings for the population for whom treatment with panitumumab is appropriate. In addition, there was a large amount of missing patient-reported outcome data that was unlikely to be missing at random: as expected for patients with advanced refractory disease, more than 50% of patients in the BSC alone arm and KRAS mutant patients in panitumumab plus BSC arm had progressed by week 8 [6], and thus missing data are likely to have been related to declining health. The last-observation-carried-forward (LOCF) method used in the previous overall population analysis of HRQoL data is therefore a less-than-optimal approach to evaluate the true impact of panitumumab on patient-reported outcomes.

Using data from this phase 3 trial, we therefore sought to evaluate the impact of panitumumab on patient-reported outcomes according to KRAS tumor status and including statistical analyses aimed at properly accounting for the missing data.

Patients and methods

Patients and study design

The patient population and design for this phase 3, open-label, randomized, controlled trial (n = 463) have been described elsewhere [6, 9]. Briefly, patients with EGFR-detectable mCRC and documented evidence of disease progression after failure of fluoropyrimidines and prespecified exposure to oxaliplatin and irinotecan were randomly assigned to receive panitumumab 6 mg/kg plus BSC every 2 weeks or BSC alone. The primary endpoint of the study was PFS, with progression assessed by central radiologic review at specified time points from weeks 8 to 48, then every 3 months thereafter. The KRAS tumor status was evaluated in a prospectively defined, retrospective analysis of formalin-fixed paraffin-embedded tumor sections in a blinded fashion, using a validated KRAS mutation kit [9]. The study protocol was approved by the ethics board at each research center, and patients provided written informed consent, including that for research into the paraffin-embedded tumor samples provided at baseline. An independent clinical research organization (working on behalf of the study sponsor) monitored study centers in Central and Eastern Europe, and representatives of the study sponsor monitored study centers in the rest of world. Monitors were responsible for reviewing adherence to the protocol, compliance with Good Clinical Practice, and the completeness, accuracy, and consistency of the data.

Patient-reported outcome instruments and data collection

Colorectal cancer symptoms were assessed using the National Comprehensive Cancer Network (NCCN) Functional Assessment of Cancer Therapy, Colorectal Symptom Index (FCSI). The FCSI is a validated, nine-item questionnaire that addresses the most important symptoms associated with CRC, including fatigue, pain, weight loss, diarrhea, nausea, stomach swelling/cramps and appetite, and more general aspects of HRQoL such as ability to enjoy life and contentedness in quality of life (QoL) [13]. Patients responded to each item of this questionnaire using a five-point Likert-type scale ranging from 0 (not at all) to 4 (very much). Raw FCSI scores, which range from 0 to 36, were converted to total scores ranging from 0 (severely symptomatic in all symptoms assessed) to 100 (symptom-free for all symptoms assessed) using published scoring methodology [13]. The minimal clinically important difference (MCID) was defined as a change in score of 3 points or more [14]. This instrument was administered at baseline and then every 2 weeks until disease progression.

Overall HRQoL was measured at baseline and monthly until disease progression using the EuroQol 5-Dimensions (EQ-5D) Index. The EQ-5D Index is a generic preference-based measure of overall QoL, as assessed across five dimensions: mobility, self-care, usual activity, pain or discomfort, and anxiety or depression. Each dimension has three possible responses (no problems, moderate problems, or extreme problems). The EQ-5D Index score is calculated by combining scores from the five dimensions using population-based preference weights and ranges from −0.594 to 1 [15]. The MCID for the EQ-5D Index in patients with mCRC has been estimated as a change in score of 0.08 points or more [16].

Statistical analysis

The primary objective of this analysis was to estimate the average difference in effect of panitumumab plus BSC treatment compared with BSC alone on the FCSI and EQ-5D Index, both overall and for the wild-type and mutant KRAS subgroups. KRAS tumor status was ascertained in 92% of patients enrolled in the phase 3 trial [9]. The analysis set was defined as all patients in the intent-to-treat population (all subjects who gave consent and were randomized) who had at least one post-baseline FCSI or EQ-5D Index assessment and an assessed KRAS status (“KRAS patient-reported outcome (PRO) analysis set”). Only outcomes obtained through week 17 of the study were used due to small sample sizes after week 17. Change in score from baseline was analyzed over time using linear mixed models for repeated measures [17]. The models included explanatory variables for study treatment arm, study week, and the interaction between treatment arm and study week. Each repeated measures model also included adjustments for baseline outcome score, baseline collection medium, baseline Eastern Cooperative Oncology Group (ECOG) performance status, and geographic region. Treatment-specific estimates of the average change in each outcome score from baseline, along with 95% CI, were calculated for the overall cohort and for each KRAS subgroup using least-squares mean (LSM) differences.

To evaluate the effect of study attrition on the estimates of treatment differences, a sensitivity analysis was performed using pattern-mixture models [18, 19], which are linear models that incorporate information about missing data. For each endpoint, patients were first categorized according to the last week for which outcome data were collected up to week 17. These categories were then collapsed into larger groups of patients representing two general patterns of missing response data: those who dropped out of the study early (“early dropout”) and those who either dropped out later or completed all assessments up to week 17 (“late dropout/completer”). For the EQ-5D Index, we defined the “early dropout” group as those who dropped out on or before week 9 and the “late dropout/completer” group as those who dropped out after week 9 or did not drop out of the study early. To account for the more frequent assessments, FCSI “early dropout” group was defined as those who dropped out before week 9 and the “late dropout/completer” group as those who dropped out on or after week 9 or did not drop out.

Dropout status was incorporated into pattern-mixture models of change in score from baseline for each outcome. These models included fixed effects for treatment arm, study week, dropout pattern group, and interactions between these effects. The model included random effects to account for variability in outcome scores at baseline and over time among patients. Within each dropout group, treatment-specific LSM estimates of average change in each outcome score from baseline, along with 95% CI, were calculated for each KRAS subgroup and overall. All statistical analyses were performed using SAS for Linux, Version 9.1 (SAS Institute).

Results

Of the 463 patients randomly assigned to treatment in the main study, information on tumor KRAS status together with post-baseline patient-reported outcome data was available for 363 patients (KRAS PRO analysis set, n = 188 panitumumab plus BSC; n = 175 BSC alone), representing 78% of the overall study population. Baseline characteristics were generally well-balanced across treatment groups and were similar to those reported in the overall patient population [6], with the exception of a slightly higher percentage of better performing patients (ECOG = 0) in the panitumumab plus BSC compared with the BSC alone treatment group, overall and for each KRAS group (Table 1).

Table 1 Baseline characteristics of randomized patients with information on tumor KRAS status plus post-baseline data on FCSI score and EQ-5D Index (KRAS PRO analysis set)

Patient-reported CRC symptoms and HRQoL

Table 2 presents the number of patients completing FCSI questionnaires for each week until week 17. In line with previous reports [20, 21], the amount of missing data was substantial, particularly in the later weeks, although patients with wild-type KRAS mCRC receiving panitumumab plus BSC had a higher percentage of available data for each post-baseline week compared with those receiving BSC alone. Similar results were seen for the EQ-5D Index (data not shown).

Table 2 Number of patients completing the FCSI (KRAS PRO analysis set)

Change in scores from baseline—linear mixed model

According to the repeated-measures model of change from baseline score, there was less deterioration in the FCSI score and EQ-5D Index in the panitumumab plus BSC group compared with the BSC alone group, both overall and for the subgroup of patients with wild-type KRAS mCRC. For the FCSI scores, an average LSM difference between treatment groups across all weeks favored panitumumab by 3.60 (95% CI, 0.90 to 6.30) overall and 5.62 (95% CI, 2.38 to 8.86) for patients with wild-type KRAS mCRC (Fig. 1a). The results of EQ-5D Index also favored panitumumab, with an average LSM difference between treatment groups across all weeks of 0.17 (95% CI, 0.09 to 0.25) overall and 0.22 (95% CI, 0.12 to 0.32) for patients with wild-type KRAS mCRC (Fig. 1b). Among patients with wild-type KRAS mCRC, in particular, the differences between treatment groups substantially exceeded each instrument’s MCID. Most notably, the difference between treatments in EQ-5D Index change was 2.75 times higher than the MCID in patients with wild-type KRAS mCRC. As expected, analysis of the mutant KRAS group did not show any significant differences in change from baseline FCSI score or EQ-5D Index between treatments (FCSI score: average LSM difference between treatment groups across all weeks of −1.41 [95% CI, −6.41 to 3.59]; EQ-5D Index: average LSM difference between treatment groups across all weeks of 0.08 [95% CI, −0.04 to 0.19]).

Fig. 1
figure 1

a Least-squares mean differences in the change from baseline in the FCSI scores between panitumumab plus BSC and BSC alone for patients with wild-type KRAS mCRC by analysis week, linear mixed models. Note: This model includes data only through week 13. Inclusion of week 17 data, which are missing for more than 80% of the analysis population, caused model instability to the extent that treatment effects could not be estimated. b Least-squares mean differences in the change from baseline in the EQ-5D Index between panitumumab plus BSC and BSC alone for patients with wild-type KRAS mCRC by analysis week (KRAS PRO analysis set), linear mixed models. FSCI Functional Assessment of Cancer Therapy, Colorectal Symptom Index; BSC best supportive care; mCRC metastatic colorectal cancer; EQ-5D EuroQol 5-Dimensions Index

Pattern-mixture model (sensitivity analysis)

For each outcome, missing data were categorized into early and late dropout patterns, regardless of treatment group or KRAS status. Based on a simple cross-sectional average, Fig. 2 illustrates that there was a clear difference in the FCSI score of the early and late dropout groups at each time point. Similar results were seen for the EQ-5D Index (data not shown).

Fig. 2
figure 2

Unadjusted average change from baseline in FCSI score by week of assessment and dropout pattern (KRAS PRO analysis set). FSCI Functional Assessment of Cancer Therapy, Colorectal Symptom Index

For both the FCSI score and EQ-5D Index, the panitumumab plus BSC group had a significantly higher proportion of patients in the late dropout group than in the early dropout group, whereas the BSC alone group had a significantly higher proportion of patients in the early dropout group (P < 0.001; Table 3). Together, these data indicate that the dropout pattern was informative about the treatment group as well as the patient-reported outcomes, and thus the missing data were considered to be non-random, and treatment effects were evaluated separately for those patients who did and those who did not drop out early.

Table 3 Summary of dropout pattern by treatment (KRAS PRO analysis set)

Table 4 displays the results of the pattern-mixture analyses modeling average differences in change from baseline score for the FCSI and EQ-5D Index. For late dropout patients with wild-type KRAS mCRC, those receiving panitumumab experienced an initial increase in EQ-5D Index and FCSI scores that was not observed in patients receiving BSC alone. In this group, the LSM difference between groups was statistically significant in favor of panitumumab with differences in both instruments exceeding the relevant MCIDs, consistent with the findings from the repeated-measures model (Table 4). On the EQ-5D index, differences between treatment groups were four times higher than the MCID and substantially exceeded the panitumumab advantage seen in the repeated-measures model. These treatment advantages were not evident for patients in the early dropout pattern group with wild-type KRAS mCRC (Table 4). Similar to the linear model, there were also no significant differences between the treatments in patients with mutant KRAS mCRC (Table 4).

Table 4 Least-squares mean differences (95% CI) in the change from baseline in the EQ-5D Index score and FCSI scores between panitumumab plus BSC vs BSC alone: overall and by tumor KRAS status, using pattern-mixture models (KRAS PRO analysis set)

Based on the FCSI assessment of symptoms among patients with wild-type KRAS mCRC who stayed on study, the treatment advantage favoring panitumumab appeared to be consistent across all weeks of the study (Fig. 3a), whereas the treatment advantage favoring panitumumab on the EQ-5D Index became larger as time increased, with the greatest advantage over BSC alone seen at week 17 (Fig. 3b).

Fig. 3
figure 3

a Least-squares mean differences in the change from baseline in the FCSI scores between panitumumab plus BSC and BSC alone for patients with wild-type KRAS mCRC in the late dropout group by analysis week, pattern-mixture model. b Least-squares mean differences in the change from baseline in the EQ-5D Index score between panitumumab plus BSC and BSC alone for patients with wild-type KRAS mCRC in the late dropout group by analysis week, pattern-mixture model. FSCI Functional Assessment of Cancer Therapy, Colorectal Symptom Index, BSC best supportive care, mCRC metastatic colorectal cancer, EQ-5D EuroQol 5-Dimensions Index

Discussion

Patient-reported outcomes such as cancer-related symptomatology and QoL are increasingly being recognized as important aspects of palliative care in the advanced cancer setting, providing important information to help make treatment decisions [12]. In this analysis, patients treated with panitumumab showed significantly better control of CRC symptoms and maintenance of HRQoL compared with those who received BSC alone: this benefit was most evident among the subgroup of patients with wild-type KRAS mCRC and was not seen in patients with mutant KRAS mCRC, or in those dropping out of the study early. This therefore illustrates, for the first time, the beneficial effects of panitumumab on HRQoL in patients with wild-type KRAS mCRC, the population for whom panitumumab treatment is indicated in clinical practice. Indeed, in those patients with wild-type KRAS mCRC treated with panitumumab who did not drop out of the study early, a decrease in CRC-related symptoms and an increase in HRQoL was observed by week 3 of the study. This is consistent with the finding that the beneficial effects of panitumumab on PFS are limited to patients with wild-type KRAS mCRC [9] and with previous analyses indicating that a lack of disease progression at week 8 provided by panitumumab was associated with better HRQoL and symptom control [20]. For HRQoL and CRC symptoms, the benefits of panitumumab compared with BSC alone in patients with wild-type KRAS mCRC exceeded the established MCID levels, indicating that statistical differences were also clinically relevant. To our knowledge, our analyses are the first to illustrate the beneficial effects of an EGFR inhibitor on CRC-specific symptomatology, which is particularly important given that treatment of patients with mCRC is primarily given with palliative intent [10, 11].

The HRQoL of patients with advanced CRC treated with the EGFR inhibitor cetuximab, overall and by KRAS status, has also been published recently [22]. The findings were generally consistent with ours in that HRQoL benefits of cetuximab were observed and were most apparent in the subgroup of patients with wild-type KRAS mCRC. However, the results cannot be more closely compared, as the cetuximab analysis used data from the EORTC QLQ-C30, which includes cancer-related (but not CRC-specific) symptoms [22], whereas we chose to evaluate HRQoL and CRC-specific symptoms separately. In addition, in the study of Au et al. [22], HRQoL data continued to be collected after disease progression. Inclusion of such data was not possible in our analysis since the study was designed to allow cross-over from BSC to panitumumab at this point [6].

Taken together with those of the cetuximab analyses [22], these HRQoL results may, however, contribute to a more accurate evaluation of the balance between the positive effects of EGFR inhibitors and the side effects associated with their use, such as skin toxicity, which are generally perceived to have a detrimental effect on HRQoL [23]. In the main study population of the phase 3 trial, skin toxicity occurred in 90% of patients in the panitumumab group compared with 9% of patients in the BSC group [6] and was more likely to develop in patients remaining on the study because of treatment benefit [21]. Despite the association of skin toxicity with prolonged treatment, the benefits of panitumumab in terms of HRQoL and symptomatology were most apparent in the group of patients with wild-type KRAS mCRC who did not drop out of the study early in our analysis. Indeed, more severe skin toxicity in association with panitumumab has been associated with higher rather than lower HRQoL scores [21]. Similar findings in relation to skin toxicity and HRQoL have also been observed with cetuximab and have been suggested to reflect an understanding by patients that skin rash is indicative of a response to therapy, which helps them cope with this side effect [24]. As noted by Au et al. [22], the presence of skin rash also introduces the potential for bias in the HRQoL results, since those patients experiencing it may have been more likely to report improvements if they were aware that the rash was a known predictor of benefit. However, this suggestion would be at odds with the fact that the benefits of panitumumab in terms of HRQoL were most prominent in patients with wild-type KRAS mCRC who were more likely to experience objective tumor control than those with KRAS mutant tumors, and these HRQoL benefits were not seen in patients with mutant KRAS mCRC despite the fact that the vast majority of this subpopulation also developed skin rash.

Our results differ from the original protocol-specified analysis of HRQoL and CRC symptom control, which did not take into account KRAS mutation status and which used the LOCF imputation method. This showed a numerical trend favoring panitumumab that did not reach statistical significance [6]. The differences in outcome highlight both the importance of tumor KRAS status and the need to properly account for missing HRQoL data in trials of late-stage disease in which the reasons for missing observations are often related to deterioration in a patient’s clinical status and so, cannot be classified as missing at random [25]. For a patient who has rapid disease progression and drops out of the study early, LOCF analysis is likely to impute values indicating better health than the true value, potentially biasing the comparison in favor of less effective or less well-tolerated treatment. An advantage to using the repeated-measures models (as in our analysis) is that imputation is not required. Pattern-mixture models provide even greater flexibility since the methodology provides a way to incorporate information about patient dropout into the longitudinal model.

Our sensitivity analyses showed that there were significantly fewer early dropouts for both the HRQoL and CRC symptom scales among the group receiving panitumumab than among the group receiving BSC alone. This disproportionate percentage of data by treatment group likely resulted from the higher rate of discontinuation due to death or disease progression in the BSC only treatment arm [6], and it was important to account for these non-random missing data. Of note is that similar findings regarding disproportionate missing data were observed in the cetuximab analysis of Au et al. [22] (although not to the same extent as there was an attempt to collect HRQoL data beyond disease progression), and similar approaches to account for missing data were included in their sensitivity analysis [22].

There are several possible explanations for the apparent lack of effect of panitumumab, in those patients (even with wild-type KRAS mCRC) who dropped out of the study early. Patients treated with panitumumab plus BSC who left the study early due to tolerability concerns or other reasons unrelated to disease progression may not have received enough treatment to yield an improvement in HRQoL or CRC symptoms. Alternatively, the treatment may not have been effective among a subgroup of patients, leading to relatively rapid progression of disease and early dropout.

A potential limitation of this study is that KRAS mutation status was determined retrospectively and therefore was not a stratification factor in the randomization. However, 92% of patients randomized had KRAS status available, and patients with KRAS mutations were well distributed between the two treatment groups. Although the post hoc nature of our exploratory analysis by KRAS status is a theoretical limitation, the comparison of baseline demographics suggests that, within the KRAS subgroups, patient characteristics were similar between the two treatment arms. Another potential limitation is the imperfect sensitivity and specificity of patient-reported outcome measures. However, our analysis used validated instruments, and our findings are supported by their consistency with the results of the primary study endpoint, PFS.

In conclusion, our analyses illustrate the clinically significant beneficial effects of panitumumab in terms of CRC-specific symptom control and maintenance of HRQoL in patients with wild-type KRAS mCRC, and these findings extend our understanding of the benefits of panitumumab treatment beyond improvements in PFS. These benefits were apparent despite skin toxicity associated with panitumumab treatment and provide important additional information for both patients and physicians when making treatment decisions for mCRC, particularly when the treatment intent is palliative.