Implementing patient-reported health-related quality-of-life data in cancer routine practice to improve accuracy of prognosis. Are we there yet?

Over the last few decades, prognostic factor analyses (PFAs) have become increasingly popular in oncology. They aim to identify ‘independent’ predictors of outcome for the patient population studied. Factors identified through such analyses are called prognostic factors, and they are often combined together into a prognostic index that will be used to classify new patients from the same population into various prognostic groups (e.g., poor-, intermediate- and good-prognosis groups). Such analyses, therefore, help to acquire increased knowledge of the disease (by identifying factors that have an impact on the outcome of the patients), help in the design of further trials or in intertrial comparisons (by identifying more homogeneous groups of patients) and, also, help in the choice of treatment for an individual patient (e.g., this information can be useful for clinicians to optimize the use of palliative care in advanced-disease settings) [1]. However, historically, PFAs in oncology have focused on patient sociodemographic characteristics, in addition to clinical and laboratory data.

In more recent years, several PFAs have also started to include patient-reported health-related quality-of-life (HRQoL) data. Generally, these studies have shown a robust link between a patient’s judgment on their own health status – measured by HRQoL questionnaires – and length of survival in advanced disease settings. This evidence has now been reproduced in a wide range of different cancer populations, including breast [2,3], colorectal [4], lung [5], melanoma [6], myeloma [7] and esophageal [8], as well as in large cohorts of patients with varied malignancies [9]. A comprehensive review [10] and a recent meta-analysis of pooled data from large, randomized, controlled trials (RCTs) [11], seem to further confirm this evidence. Some studies have shown that patients’ self-reported HRQoL parameters not only independently predict survival but also predict survival better than important clinical data. There are a number of studies, for example, where the final multivariate model predicting survival retained HRQoL parameters in place of important clinical data, such as performance status or tumor size [12–14]. As an example, while performance status has been considered an important prognostic variable for many years, and has been used extensively as a stratification factor in RCTs, several studies in palliative-care settings have shown that patients’ HRQoL parameters have a stronger prognostic value than performance status [7]. From an intuitive point of view, this would make great sense, as it might be assumed that a patient’s evaluation of their own health status might, indeed, provide more accurate information than what could be inferred by asking someone else (i.e., the physician).

The link between HRQoL parameters and survival is a new important point to consider in cancer research, since, in effect, it points out the fact that information given by patients to clinicians through the use of standardized reports (i.e., questionnaires) provides clinically meaningful information about prognosis that is far beyond what we have known for many years to be prognostic for cancer patients. Potentially, this finding could have crucial clinical implications, for example, by allowing clinicians using HRQoL information in routine practice to have additional prognostic information on that particular patient. In this scenario, which is ideally one of the main goals of this line of research, the physician could simply ask the patient to fill in a given questionnaire and then interpret the score for prognostic purposes. This way, the physician would have an additional tool in his hands to possibly help him make more-tailored treatment decisions.

Considering the large number of studies published in this area to date, the majority of which find at least one HRQoL parameter to be an independent prognostic variable, a critical question is how to move this area of research forward. How long does it take to implement this evidence into clinical practice? Are we not already there?

We must note that PFAs that include HRQoL data have often reported inconsistent results regarding which HRQoL parameter is prognostic for survival in a given cancer population. Considering, for example, studies conducted in patients with advanced non-small-cell lung cancer, various HRQoL parameters have been identified (even when using the same HRQoL questionnaire). For example, using the European Organisation for Research and Treatment of Cancer (EORTC) Quality-of-Life Questionnaire (QLQ)-C30, Herndon et al. found pain in their multivariate analysis [15]; Montazeri et al. identified that the global HRQoL scale was prognostic [5] and, more recently, Efficace et al. found that both pain and dysphagia independently predicted overall survival [16].

Two main methodological points are worth emphasizing here, since they might explain this lack of consistency: first, there is an impressive heterogeneity of study designs and statistical methods used and, second, very few of these studies have validated their results on an independent dataset of patients; therefore, less confidence can then be given to the results.

The studies in this field have used different HRQoL questionnaires and different statistical approaches and, in the large majority, HRQoL analyses were conducted on an exploratory basis without being described a priori in the protocol. Other issues are also related to the time of the baseline HRQoL assessment and the related time since diagnosis. In some studies, for example, the baseline assessment was obtained by surveying patients within 2 years of diagnosis [17]. In addition, decisions to be taken in the PFA of the data, such as the a priori selection of factors to be studied in regression analyses, choice of cutoff values for continuous variables and statistical model building strategies, may lead to different results. In this journal, Expert Review of Pharmacoeconomics and Outcomes Research, Mauer and colleagues reported a detailed and comprehensive review of research methods employed, highlighting a number of statistical issues deserving more consideration in future studies [18]. Similar to what occurs with medical data, conflicting results can be due to flaws in statistical methodologies used to analyze the data, and some of the difficulties encountered in the PFA of clinical data seem to be even more problematic when considering HRQoL data. Multicollinearity is a well-known problem in PFAs, and is even more problematic when considering HRQoL data. PFAs are based on multiple regression models in which multicollinearity occurs when two or more predictor variables are so highly correlated that unstable results are obtained. Multicollinearity can either lead to incorrect model selection or, when the correct model specification is nevertheless achieved, to incorrect direction (sign) and/or magnitude of effects of the predictor variable on the response variable [19]. This makes the results obtained very difficult to interpret, and these issues clearly have to be addressed before conclusions can be drawn from a PFA. While multicollinearity can often be avoided when considering clinical variables by appropriate preselection of the variables chosen for the analysis, risk of harmful multicollinearity has to be carefully evaluated when adding HRQoL factors into PFA, as multicollinearity is often inherent to the questionnaire itself, all variables being designed to measure putative components of HRQoL [20]. In addition, HRQoL questionnaires used in oncology studies are all multiple items, and the greater the number of items, the higher the risk of multicollinearity. Van Steen et al. discuss diagnostic tests and their limitations and propose proper identification tools for the impact of multicollinearity in PFAs using HRQoL variables [20]. Following a bootstrap resampling technique proposed by Sauerbrei et al.[21], they developed, in the context of QoL, a bootstrap model averaging technique, which offers a better insight into the stability of the final model predicting survival.

As conflicting results are not uncommon in PFAs, strong recommendations have been made to check the reliability and reproducibility of results obtained. ‘Overfitting’ is, indeed, a well-known phenomenon in PFAs, when the model obtained is too specific to the data used to develop it. There are several reasons why a PFA can lead to different results on different samples and the ‘validation of prognostic index’, preferably on an independent sample, is a major component of PFA [22,23]. Recently, Efficace and colleagues highlighted that patient self-reported social functioning, as measured by the EORTC QLQ-C30, is an independent prognostic factor for survival in metastatic colorectal cancer patients, and they also validated this finding in an independent population [24,25]. However, validation of results in this area is still rare, and more independent validation studies are needed in other cancer disease sites.

In addition to the aforementioned methodological challenges, it is also worthy of note that the bulk of this evidence mainly stems from patients enrolled in RCTs, which usually include patients with selected entry criteria. If we are to eventually implement this evidence into clinical practice, it would be of value to rely on data from observational studies, which are currently lacking. Only in an observational setting would it be possible, for example, to take into account variables that are unlikely to be reported in a RCT of drug efficacy. Results from observational study designs could help strengthen the evidence of a link between HRQoL parameters and survival duration in the real world, for example, by taking into account older patients and those with comorbidity.

Another interesting question is how to explain the reasons of association between HRQoL data and outcome. Indeed, after having shown that a specific HRQoL aspect provides a good and strong indicator of the patient’s prognosis independent of other known biomedical factors, reasons underlying this association may be of interest. Does the HRQoL factor identified as a prognostic factor act as a proxy for other biomedical factors not yet identified as prognostic factors (e.g., the patient feels better because he actually is better) or does the wellbeing of the patient act on the outcome of the patient? Both of these discussion points require the availability of good data and good techniques of analyses.

In conclusion, while it is true that results in this area are not yet consistent, and methodological issues still need to be resolved, it is worth noting that we are still confronted with a relatively new line of research, as the majority of studies were published over the last decade [10]; thus, current results are to be considered a work in progress. Nevertheless, the amount of evidence obtained to date in several advanced cancer populations is convincing enough to state that patient-reported HRQoL data, collected through standardized questionnaires, play a key role in revealing prognostic information. The way this information can then be concretely implemented into future prognostic indices and fit into the routine oncology practice is, indeed, an additional key topic for the next HRQoL research agenda.

Financial & competing interests disclosure

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

No writing assistance was utilized in the production of this manuscript.