Longitudinal health-related quality of life assessment: implications for prognosis in ovarian cancer

We examined 137 histologically confirmed ovarian cancer patients treated at Cancer Treatment Centers of America® at Midwestern (MRMC) and Southwestern (SRMC) Regional Medical Centers between Jan 2001 and Dec 2009 who had QoL data available at both baseline and 3-month follow-up. The inclusion criteria for participation in this study were a histological diagnosis of ovarian cancer and the ability to read English. Patients with all stages of ovarian cancer were eligible for the study. Patients were excluded if they were unable to give informed consent or were unable to understand or cooperate with study conditions.

A trained clinical coordinator was responsible for determining eligibility, describing the study, and obtaining informed consent. All patients were assured that refusal to participate would not affect their future care in any way. Patients who chose to participate were presented with the questionnaire at their initial visit and instructed to return their completed questionnaires to the clinical coordinator within 24 hours. Thus, patients completed baseline questionnaires prior to receiving therapy at our facility. Following the completion of the baseline questionnaire, all patients were treated with an integrative model combining surgery, radiation and chemotherapy as appropriate, plus complementary therapy consisting primarily of nutritional, psychosocial, and spiritual support, naturopathic supplements, pain management, and physical therapy/rehabilitation. Patients completed the questionnaire again after a follow-up period of 3 months.

Additional patient data recorded for this study was age at presentation (current age), stage of disease at diagnosis and prior treatment history (previously treated versus newly diagnosed). The only follow-up information required was the date of death or the date of last contact/last known to be alive, obtained from the tumor registries at Midwestern and Southwestern Regional Medical Centers. This study was approved by the Institutional Review Board at Cancer Treatment Centers of America®.

QoL was assessed using the European Organization for the Research and Treatment of Cancer Quality of Life Questionnaire (QLQ-C30), which emphasizes a patient’s capacity to fulfill the activities of daily living. The QLQ-C30 is a 30-item cancer specific questionnaire that incorporates five functioning scales (physical, role, cognition, emotional, and social), eight symptom scales (fatigue, pain, and nausea/vomiting, dyspnea, insomnia, loss of appetite, constipation, diarrhea), financial well-being scale and a global scale (based on two items: “how would you rate your overall health during the past week” and “how would you rate your overall quality of life during the past week”). The raw scores are linearly transformed to give standard scores in the range of 0–100 for each of the functioning and symptom scales. Higher scores in the global and functioning scales and lower scores in the symptom scales indicate better QoL. A difference of 5–10 points in the scores represents a small change, 10–20 points a moderate change and greater than 20 points a large, clinically significant change from the patient’s perspective [17]. This instrument has been extensively tested for reliability and validity [18‐20].

Patient survival was the primary end point and defined as the time interval between the date of first patient visit to the hospital and the date of death from any cause or the date of last contact/last known to be alive. Two separate analyses were performed. First, the relationship between baseline QoL and patient survival was investigated for 137 patients. Second, the relationship between change in QoL scores between baseline and 3 months and survival was assessed for the same patient cohort. Change scores were calculated by subtracting the baseline scores from the 3-month QoL scores. The overall survival was calculated using the Kaplan-Meier method. Clinical and QoL variables were evaluated using univariate Cox proportional hazards models to determine which parameters showed individual prognostic value for survival. Multivariate Cox proportional hazards models were then performed to evaluate the joint prognostic significance of all QoL and clinical factors. Each QLQ-C30 scale was treated as a continuous variable for the purpose of Cox regression analyses. The effect of QoL parameters on patient survival was expressed as hazard ratios (HRs) with 95% confidence intervals (CIs). Changes of 10 or more points on a 0 to100 scale are considered clinically relevant [21], so we present hazard ratios for a 10-point change on the continuous QoL variables. An effect was considered to be statistically significant if the p value was less than or equal to 0.05. All statistical tests were two sided. All data were analyzed using IBM SPSS version 20.0 (Chicago, IL, USA).

Cox regression with time-invariant covariates assumes that the ratio of hazards for any two groups remains constant in proportion over time. We checked this assumption by first examining log-minus-log plots for the categorical predictors and then fitting a Cox regression with a time-varying covariate for each predictor in turn. Potential multicollinearity was assessed in two steps. Large values (above 0.75) of Pearson’s correlation coefficients were used as an initial screen for pairs of QoL variables. As a second check, the variance inflation factor was used with the final model to verify that multicollinearity was not significantly influencing model coefficients [22, 23]. In order to minimize instability of the final multivariate model resulting from high multicollinearity, global QoL was evaluated separately because it is most highly correlated with all other variables on the QLQ-C30 questionnaire, and also because it is difficult to interpret and manipulate clinically. Finally, to assess the possible influence of sample bias on the results, as well as to investigate the stability of the model coefficients, we performed a bias-corrected and accelerated (BCa) bootstrap resampling procedure. We generated 1000 samples, each the same size as the original data set, by random selection with replacement. Cox regression was then run separately on these 1000 samples to obtain robust estimates of the standard errors of coefficients, and hence the p values and 95% BCa CIs of the model coefficients [24].

Table 1 describes the baseline characteristics of our patient cohort. All newly diagnosed patients received appropriate de-bulking surgery followed by a platinum-based chemotherapy. All patients with recurrent disease received systemic chemotherapy chosen by their attending physician based on prior treatment history. Table 2 displays the baseline QoL scores across the two categories of prior treatment history. There were no statistically significant differences in QoL scores between newly diagnosed and previously treated patients.

Table 3 describes the results of univariate Cox regression analysis for baseline patient characteristics. Treatment history was significantly associated with survival while age and stage at diagnosis were not. Median overall survival for the entire patient cohort was 33.5 months (95% CI: 11.5-55.6 months). The median survival for newly diagnosed and previously treated disease was 66.4 and 19.6 months respectively, log-rank p<0.001, which is not surprising given that previously treated patients had advanced stage/recurrent disease at the time of presentation to our institution.

Table 4 describes the baseline scores for all dimensions of EORTC QLQ-C30 instrument. Among the EORTC QLQ-C30 functioning scales, social functioning had the lowest (worst) mean score of 66.5 while the highest (best) mean score of 78.3 was recorded for cognitive functioning. Among the EORTC QLQ-C30 symptom scales, nausea/vomiting had the lowest (best) mean score of 13.1 while the highest (worst) mean score of 40.0 was recorded for fatigue. Table 4 also displays the results of univariate and multivariate Cox regression analyses for each QoL variable. The HRs along with their 95% CIs for every 10-point increase in all EORTC QLQ-C30 scales are given. On univariate analysis, no baseline QoL variables were predictive of survival. Before proceeding with multivariate analysis, we checked the bivariate Pearson’s correlation among the QoL variables to screen for observable multicollinearity. All correlation coefficients were smaller than the pre-decided cut-off level of r=0.75. As a result, all QoL variables were considered in the multivariate analysis. On multivariate analysis, no QoL function or symptom variable was found to be statistically significantly associated with survival. A separate multivariate model was run for global QoL after adjusting for prior treatment history. Global QoL was not found to be significantly associated with survival. VIF values for baseline QoL variables ranged from 1.1 (diarrhea) to 4.3 (fatigue), none of which indicates a significant problem with multicollinearity [22, 23]. There was no evidence of non-proportional hazards in the multivariate models presented.

In order to further investigate the stability of the classical multivariate Cox models reported in Table 4, we conducted a bootstrap resampling procedure based on 1000 samples. The bootstrap estimates of the multivariate HRs along with corresponding p values and 95% BCa CIs are provided in Table 5. We found no significant differences in regression coefficients and their corresponding p values between the classical Cox regression and bootstrap Cox regression models.

Table 6 describes the change in scores from baseline to 3 months for all dimensions of EORTC QLQ-C30 instrument. On average, they were small. Table 6 also displays the results of univariate and multivariate Cox regression analyses for change in QoL scores. On univariate analysis, the only QoL scale that was significantly predictive of survival was global function. Before proceeding with multivariate analysis, we checked the bivariate Pearson’s correlation among the change scores to screen for observable multicollinearity. All correlation coefficients were smaller than the pre-decided cut-off level of r=0.75. As a result, all QoL change variables were considered in the multivariate analysis. On multivariate analysis, change variables that were significantly predictive of survival were appetite loss and constipation. The corresponding HRs for 10-point increase (deterioration) in appetite loss and constipation scales were 1.20 (95% CI: 1.06 to 1.35; p=0.005) and 1.13 (95% CI: 1.02 to 1.24; p=0.02) respectively. A separate multivariate model was run for change in global QoL after adjusting for prior treatment history. Every 10-point increase (improvement) in global function from baseline to 3 months was associated with a 10% decreased risk of death (HR=0.90; 95% CI: 0.81 to 0.99, p=0.03). VIF values for change in QoL variables ranged from 1.2 (change in diarrhea) to 3.9 (change in fatigue), none of which indicates a significant problem with multicollinearity. There was no evidence of non-proportional hazards in the multivariate models presented.

In order to further investigate the stability of the classical multivariate Cox models reported in Table 6, we conducted a bootstrap resampling procedure based on 1000 samples. We found no significant differences in regression coefficients and their corresponding p values between the classical Cox regression and bootstrap Cox regression models except for the constipation symptom which changed from being statistically significant in the classical model to being marginally significant in the bootstrap model (results not shown in the interest of space).

We decided to explore the data further by conducting separate multivariate analysis for previously treated patients. Similar to what we observed in the total sample above, improvement in appetite, constipation and global health scores during the first 3 months of treatment was significantly associated with improved survival time in previously treated patients.