Background

Colorectal cancer (CRC) is the third most diagnosed cancer in men and the second most diagnosed in women, with over 1.2 million new cases and 608 700 deaths worldwide in 2008[1]. About up to half (20% to 50%) of CRC patients will develop metastases during the course of their disease[2] and approximately 35% are diagnosed with synchronous metastases[2, 3]. Standard treatments for metastatic CRC (mCRC) are based on chemotherapy.

As is the case for many cancers, CRC staging is essential for optimal patient management. Accurate prognostication facilitates both therapeutic decisions and stratification in randomized clinical trials of cancer treatments. In CRC, the well-known TNM staging system is predominantly used[4]. In mCRC, two validated prognostic classification systems can be applied: Köhne prognostic index[5] for patients receiving front-line fluoropyrimidine mono-chemotherapy and GERCOR (Groupe Coopérateur Multidisciplinaire en Oncologie) prognostic index[6] for patients with oxaliplatin-based or irinotecan-based regimens. However, the models’ ability to discriminate between patients on the basis of their prognosis (as measured by the C-index[7]) is still relatively modest. Thus, improvement of these prognostic indicators is required[6].

In palliative care patients, the prognostic value of health-related quality of life (QoL) has been demonstrated for several types of cancer[8–10]. For mCRC patients, QoL is known to be an independent prognostic factor for overall survival (OS)[8, 11]. Hence, QoL is a candidate for the improvement of existing prognostic indices. Given that QoL is a multidimensional concept, there is a need to identify the QoL dimensions associated with OS for each specific type of cancer. The results of a recent study showed that social functioning (as measured with the EORTC QLQ-C30 tool) is an independent prognostic factor for survival in mCRC patients[12]. The objective of the present study was to assess the independent prognostic value of QoL in mCRC and evaluate its ability to improve the Köhne and GERCOR prognostic indices.

Methods

Patients

Individual patient data from the OPTIMOX1 phase III trial were analysed. The 620 evaluable patients from OPTIMOX1 were recruited from January 2000 to June 2002 from 56 institutions in five countries. In this trial, the oxaliplatin stop-and-go strategy proved to be as good as a continuous oxaliplatin-based chemotherapy strategy in previously untreated mCRC patients. The trial's inclusion and exclusion criteria are detailed elsewhere[13].

Quality of life assessment

Quality of life was self-reported by the patient using the generic EQ-5D questionnaire (also known as EuroQol)[14], which has five dimensions (mobility, self-care, usual activities, pain/discomfort and anxiety/depression) rated as one of three levels ("no problems",”some problems" and ”extreme problems", coded as 1, 2 and 3, respectively). The EQ-5D also includes a 100-centimetre visual analogue scale (VAS) for the self-assessment of overall health (0 = worst possible score; 100 = best possible score).

The GERCOR and Köhne prognostic indices

The Köhne prognostic index[5] comprises four variables: performance status (PS), number of metastatic sites, alkaline phosphatase (ALP) level and white blood cell (WBC) count. The GERCOR prognostic index[6] is based on two variables: PS and serum lactate dehydrogenase (LDH) level. Patients are classified into three risk groups (low, intermediate and high) in both models.

Statistical analysis

Demographic and clinical characteristics were summarized as frequency and percentage. In order to check whether selection bias was present, the patients’clinical characteristics were compared (with chi-squared test or Fisher's exact test) as a function of the available QoL data at baseline.

Overall survival was defined as the time from randomization to death (regardless of the cause) or last follow-up (censored data). All randomized patients with complete QoL data were included in the statistical analysis.

Univariate and multivariate analysis were performed using Cox proportional hazards modelling, with calculation of the hazard ratio (HR) and the corresponding 95% two-sided confidence intervals (95%CI).

In order to evaluate the independent prognostic value of QoL, we built two multivariate models with backward selection. The first model included all demographic and clinical variables associated with OS (p<0.1) in univariate analysis. The second model included demographic, clinical and QoL variables with p<0.1 in univariate analysis.

Improvements in the prognostic index was evaluated by adding clinical variables (other than those used to build the prognostic index) and QoL variables (with p<0.1 in univariate analysis) to a model with backward selection (Köhne or GERCOR index being forced in the model). Patients with available QoL data for whom Köhne and GERCOR indices could be calculated were considered for prognostic systems’ improvement.

The models were compared by calculating the Schemper statistic[15] and Harrell’s C index[7]. The Schemper statistic is equivalent to R2 in linear regression and quantifies the proportion of the survival variability that is explained by the model. Briefly, the higher the Schemper statistic is, the more accurate the OS predictions would be. Harrell’s C index estimates discriminate capability, i.e. the ability to distinguish between high-risk and low-risk patients. The C-index varies from 0.5 (no discrimination) to 1 (perfect discrimination). Optimism-corrected C-index was calculated using 200 bootstrap replications.

Category-free net reclassification improvement[16] (NRI) was also calculated at various moments (12, 24 and 36 months), in order to evaluate the additional utility of QoL domains and other clinical factors. NRI quantifies”the correctness of upward and downward reclassification or movement of predicted probabilities as a result of adding a new marker”. The confidence interval for NRI was calculated using the percentiles of 1000 bootstrap replications.

We also performed a sensitivity analysis using the multiple-imputation technique[17, 18] (with 10 replications) for missing QoL data. The choice of 10 replications was prompted by the large amount of missing QoL data in the trial (60%). In line with Van Buuren’s method[19], the demographic and clinical variables initially included in the final complete-data model, those associated with the lack of QoL data and those strongly associated with OS (albeit absent from the final model) were used as predictors for the imputation of missing QoL data using a logistic regression model (QoL coded as 2–3 vs. 1). Multiple imputation with 10 replications (of the original database) consisted in creating 10 plausible values for each missing data and thus generating 10 new complete databases. For each of the new databases, a standard analysis was performed and the results were combined into a single estimation of the parameter of interest, while taking account of the uncertainty of the imputation technique[20]. Variables selected more than 5 times out of 10 replications were included in the multivariate model after multiple imputations.

Since there was no within-imputation variance according to the Schemper statistic, the pooled estimate was presented as the median [range][20].

Construction of the a modified prognostic index was based on linear transformation as follows: The regression coefficient for each variable selected in the final multivariate complete case Cox model was divided by the lowest coefficient and rounded to the nearest integer[21]. The sum of these integers is the maximum score (M) for the modified index; hence the new score varied from zero to M. According to the score, the modified prognostic index was then arbitrary divided into three risk groups: good prognostic, intermediate prognostic and poor prognostic.

Survival distributions were estimated using the Kaplan-Meier method[22] and compared with the log-rank test.

All statistical analyses were carried out using SAS® software (version 9.2, SAS Institute Inc., Cary, NC) and R.2.12.0 software (free) using the Design, SurvIDINRI (for NRIs) and Multivariate Imputation by Chained Equations packages (http://www.multiple-imputation.com/). P-values were two-sided and variables with p<0.05 were considered significantly associated with OS in multivariate models.

Results

Patient characteristics

The patient baseline characteristics are summarized in Table 1, most of them were male (59%) and 43% were over the age of 65. Synchronous metastasis was predominant (68%) and most of the patients with metachronous metastasis received adjuvant chemotherapy (66%, 130/196).

Table 1 Baseline demographic, clinical and laboratory variables for patients with and without available QoL data
Full size table

Data on QoL was available for 249 of the 620 patients in the original OPTIMOX1 cohort (40%). Normal serum LDH was significantly more frequent in patients with missing QoL data. Patients with missing QoL data also tended to have lower serum ALP levels, a better PS and less liver involvement compared to patients with available QoL. Of the 249 patients, 75% died after a median follow-up period of 35.8 months (95% CI = [33.8–38.4]). There was no apparent correlation between the availability of QoL datasets and OS (Log-rank pvalue = 0.62; Figure 1).

Figure 1
figure 1

Overall survival (in months) of patients lacking QoL data (dotted line; n = 371) and patients with available QoL data (solid line; n = 249). Log-rank p value = 0.62. The median survival times for patient with and without QoL datasets were 18.6 months (95% CI [17.0 - 21.6]) and 20.8 months (95% CI = [19.5–22.2]), respectively.

Full size image

Most of the patients had good QoL: 81%, 93%, 71%, 50% and 53% had no problems in terms of mobility, self-care, usual activities, pain/discomfort and anxiety/depression, respectively. The median VAS score was 70 (range = [10–100]).

Univariate analysis

Given that”extreme problems” (coded as 3) were infrequent, QoL item scores were pooled into two classes (i.e. a score of 1 vs. a score of 2 or 3). We also combined PS into 2 classes (0 vs. 1–2), due to the low proportion of patients with a PS score of 2.

Univariate analyses of clinical and QoL variables are summarized in Table 2. High serum LDH, poor PS, high serum ALP, >1 metastatic sites, age>65, high serum CEA, mobility problems (as coded 2–3) (Figure 2), pain/discomfort problems (as coded 2–3) and anxiety/depression problems (as coded 2–3) were associated with a poorer prognosis.

Table 2 Univariate and multivariate Cox analyses
Full size table
Figure 2
figure 2

Overall survival (in months) of patients with mobility problems (as coded 2–3) (dotted line; n = 54) and patients without mobility problems (as coded 1) (solid line; n = 223). Log-rank p value = 0.0011. The median survival times were 20.9 (95% CI = [18.6–24.9]) months and 11.8 (95% CI = [11.1–17.3]) months for patients without problems (coded as 1) and those with problems (as coded 2–3), respectively.

Full size image

There were no significant associations between the risk of death and self-care (p = 0.1322), usual activities (p = 0.2553) and the VAS score (p = 0.1280) QoL scales on the other.

Multivariate analysis

The results for multivariate analyses are summarized on Table 2.

In the first model, high LDH, >1 metastatic sites and poor PS were associated with a shorter survival.

In the second model, high LDH, >1 metastatic sites, poor PS and mobility problems were associated with a shorter survival.

After multiple imputations, the pooled HR for mobility was 1.57 (95% CI = [1.16–2.12]) (p = 0.0043) in the model including LDH, the number of metastatic sites, PS, ALP, pain/discomfort and mobility (Additional file1).

Improvement of prognostic indices

In order to evaluate improvements in performance of the Köhne and GERCOR prognostic indices, we first calculated their performance in our population (Table 3).

Table 3 Improvement of Köhne prognostic index
Full size table

Improvement of the Köhne prognostic index

After addition of QoL and clinical variables to the Köhne prognostic index in a complete-case analysis (N = 236), high LDH, mobility and pain/discomfort problems appeared to be related to a shorter survival (Table 4). The C-index and Schemper statistic were improved while the NRIs were significantly different from zero (Table 3). A modified Köhne prognostic index was built using the above variables (Table 5).

Table 4 Improvement of the GERCOR prognostic index
Full size table
Table 5 Modified Köhne prognostic index
Full size table

Survival distributions for the Köhne and improved Köhne prognostic systems are shown in Figure 3 A&3B.

Figure 3
figure 3

Survival strata according to the Köhne prognostic model before and after improvement. A: Overall survival (in months) for good, intermediate and poor prognosis according to the Köhne prognostic model. Median survival = 20.7 [17.7 – 24.4] for the group with good prognosis (n = 134); Median survival = 18.6 [17.1 – 25.4] for the group with intermediate prognosis (n = 84); Median survival = 9.0 [7.3 -14.7] for the group with poor prognosis (n = 18). Log-rank p = 0.0013. Optimism corrected C-index = 0.54. B: Overall survival (in months) for good, intermediate and poor prognosis according to the modified Köhne group. Median survival = 27.0 [21.1 – 37.5] for the group with good prognosis (n = 57); Median survival = 18.4 [16.5 – 21.6] for the group with intermediate prognosis (n = 146); Median survival = 11.3 [9.0 – 16.9] for the group with poor prognosis (n = 33). Log-rank p<0.0001. Optimism corrected C-index = 0.60.

Full size image

The Results of multiple imputations are summarized in Table 3.

A complete-case analysis of the GERCOR prognostic classification revealed that mobility and Anxiety/depression could improve performance: the C-index, Schemper statistic, and NRI are summarized in Table 4.

Based on these two new QoL scales, a modified GERCOR prognostic system was built using the above variables (Table 6).

Table 6 Modified GERCOR prognostic index
Full size table

Survival distributions for the GERCOR and improved GERCOR prognostic systems are shown in Figure 4A and Figure4B.

Figure 4
figure 4

Survival strata according to the GERCOR prognostic model before and after improvement. A: Overall survival (in months) for good, intermediate and poor prognosis according to the GERCOR prognostic system. Median survival = 28.7 [24.5 – 38.9] for the group with good prognosis (n = 73); Median survival = 19.9 [18.1 – 23.9] for the group with intermediate prognosis (n = 97); Median survival = 12.1 [10.0 – 15.4] for the group with poor prognosis (n = 66). Log-rank p<0.0001. Optimism corrected C-index = 0.65. B: Overall survival (in months) for good, intermediate and poor prognosis according to the modified GERCOR prognostic system. Median survival = 28.2 [24.5 – 37.5] for the group with good prognosis (n = 68); Median survival = 21.6 [18.7 – 26.2] for the group with intermediate prognosis (n = 90); Median survival = 11.5 [10.0 – 14.7] for the group with poor prognosis (n = 78). Log-rank p<0.0001. Optimism corrected C-index = 0.66.

Full size image

The Results of multiple imputations are summarized in Table 4.

Discussion

In this study, EuroQol mobility dimension appeared to be the third most important prognostic factor (measured by the hazard ratio) for overall survival in unresectable mCRC, after serum LDH level and ECOG performance status. Self-reported QoL is known to be associated with OS in several types of cancer[8, 9, 11, 12]. Our present results confirmed the independent prognostic value of QoL scales in patients with mCRC[8, 11, 12]. Our first multivariate model (including clinical and biochemical variables) revealed the prognostic value of LDH, PS and the number of metastatic sites, whereas our second model (with the addition of QoL) confirmed the value of LDH, PS and the number of metastatic sites and further identified the QoL”mobility" scale as an independent prognostic factor.

After multiple imputations, the mobility QoL scale remained significant despite its high associated relative increase in variance due to missing data imputation. Pain/discomfort was not significant but showed a prognostic value after the multiple- imputation analysis; this may be partially related to the high increase in variance due to missing QoL data.

We found that the Köhne prognostic system could be improved by including LDH, mobility and pain/discomfort in both complete-case and imputation analyses. Moreover, the GERCOR prognostic index was improved by mobility and anxiety/depression in a complete-case analysis and by ALP, mobility and pain/discomfort after multiple imputations. This difference in the selection of variables may be due to lack of power in the complete-case analysis albeit ALP was at the limit of statistical significance. Therefore the GERCOR prognostic index was essentially improved by QoL scales. The added value of QoL scales (completed by the patient) for improvements of the two prognostic systems revealed that the patient’s perception of his/her disease was an important information to record for prognosis assessment in addition to the clinician’s evaluation[23].

Despite a marked increase in variance due to missing data, the mobility and pain/discomfort QoL dimensions significantly improved the Köhne and GERCOR staging systems. This result comforted the independent prognostic value of these QoL scales in mCRC patients. The results for complete-case and multiple-imputation analysis were very similar. QoL significantly improved the prognostic indices with both methods (complete-case and multiple-imputation analyses). This may be related to the fact that the compete-case analysis was not biased. In fact, patients with and without QoL data at inclusion did not differ in terms of the median survival time[24] (i.e. missingness was not related to outcome).

It should be noted that such a large improvement in the C-index from 0.54 to 0.66 for the Köhne prognostic index has rarely been reported in prognostic studies. After the addition of both clinical and QoL factors, the NRIs were also statistically significant for both the Köhne and the GERCOR prognostic systems (95% CIs did not contained zero). The independent prognostic value of mobility and pain/discomfort QoL scales (using the EQ-5D) for mCRC is compatible with the result of Efficace[12] regarding the prognostic value of social functioning scale (using the EORTC QLQ-C30). In fact, mobility and pain problems could impair the social functioning QoL dimension.

One of the present study's strengths relates to its use of the easily understood and rapidly completed EQ-5D. The EQ-5D was chosen because it was expected to be less time consuming and could prevent missing data. However, EQ-5D is not a cancer-specific questionnaire like the EORTC QLQ-C30 and it constitutes a limitation of our study. The high proportion of missing data (60%) and its large variability between countries (ranged from 5% to 66%) constitute another limitation in the generalizability of our results. Such a large heterogeneity in missing data might be related to the trial logistic and/or each country’s culture. It is also important to note than our population came from a randomized controlled trial with restrictive inclusion and non inclusion criteria and might not be representative of mCRC patients in general[25]. Quality of Life may be an important parameter to record when assessing the situation of mCRC patients, since it improved the accuracy of OS prediction and greatly improved the two best-known prognostic classification systems for mCRC. We consider that QoL domains are important factors in the field of stratified therapy in the sense that knowing some aspect of the patient’s self-reported QoL level could be decisive in the choice of different treatment options in the area of tailored medicine. By way of an example, a clinician might wish to avoid a treatment with pain as side-effect if the patient reported preexisting pain symptoms. Pain and mobility could also serve as an inclusion and/or stratification factor in randomized, controlled trials in mCRC.

Conclusion

Our results confirmed the prognostic value of QoL in mCRC patients. Thus, QoL scores should be recorded as it could give supplementary information to the clinician regarding the prognosis of a patient as well as in the judgment of an acceptable treatment side effect.

Acknowledgements

The authors thank Dr. David Fraser for advice in English language.