For the evaluation of cancer treatments, parameters such as survival time or time to tumor progression are the most common primary outcomes. In addition, assessment of patient-reported physical or psychosocial symptom burden has gained considerable importance within clinical trials [1–4] and clinical practice [5]. Patient-reported outcomes (PROs) can add important information to clinical trials and medical decision making [6]. To improve the interpretation of PRO data collected within clinical trials, research efforts have focused on thresholds like minimally important differences (MIDs) or minimally important changes (MICs) over time. An MID or MIC is the smallest difference or change in PRO score points considered to be of clinical relevance [7]. Such thresholds related to change scores help to interpret the impact of disease and treatment in a relative manner by comparison to previous assessments of the same patient or to other patient groups. However, they do not aid in the interpretation of absolute scores, i.e. scores obtained from an individual patient or a patient group at a single time point. To better understand what a given score from an individual patient or an average score from a patient group means, researchers have started to develop clinical thresholds that allow identification of clinically important symptom burden or impairment [8, 9].

The EORTC QLQ-C30 [10] is one of the most widely used health-related quality of life (HRQOL) questionnaires in cancer research. It assesses important functioning domains (e.g. physical, emotional, role) and common cancer symptoms (e.g. fatigue, pain, nausea/vomiting, appetite loss). Current approaches to defining clinical thresholds for the EORTC QLQ-C30 vary substantially and each has its own strengths and weaknesses. The simplest and most straight-forward approach is to rely on the wording of the item or response categories themselves, and classify a patient as having a clinically important problem if s/he responds with at least “a little” for any given item or domain assessed on the 4 point response scale (i.e. “not at all,” “a little,” “quite a bit,” and “very much”) [11, 12]. This approach can be problematic as it uses the same threshold for all QLQ-C30 domains. As has been shown in Item Response Theory analyses [13], the QLQ-C30 items differ in item difficulty (i.e. they assess different levels of severity of a problem or symptom), suggesting that scores are not directly comparable across domains. For example, reporting “quite a bit” of vomiting probably indicates a different symptom level than reporting “quite a bit” of trouble with sleeping. This problem also applies to studies using the same cut-off score as a threshold for all domains (e.g. a score of 50 as done by Klinkhammer et al. [14]).

Another approach is based on score distributions and makes use of percentiles or statistics from reference populations (often the general population). However, the choice of a percentile or statistic is somewhat arbitrary. In addition, if, as is commonly done [15–17], the general population mean is used as a threshold, the resulting percentage of patients with a clinically important problem is likely to be very high. Using the population average often means that around half of the general population (depending on the score distributions) will be classified as having a clinically important problem. Also, such a definition of clinical thresholds assumes the prevalence rate of clinically important problems is constant across the various HRQOL domains.

These limitations can potentially be overcome by relying on external anchors for determining thresholds for clinical importance. An example of this approach is the use of patients’ ratings of their most bothersome symptoms or functional limitations as an external anchor. Patients could be asked to name their most bothersome symptoms or limitations in addition to completing scales assessing symptom or problem levels. This allows one to determine thresholds above which patients are likely to consider a symptom or limitation as bothersome (e.g. using ROC analysis). Although this has clear advantages over distribution-based thresholds, it relies on a somewhat narrow definition of clinical importance (basically being equated with what is most “bothersome”). Also, in previous work the number of most bothersome symptoms per patient has been fixed a priori [18].

In the current study, we have investigated an alternative approach to establishing thresholds for clinical significance that does not rely on one, but rather on several potential anchors. The objective of our study was to use an anchor-based approach to defining thresholds for clinical importance for four key EORTC QLQ-C30 scales: Physical Functioning (PF), Emotional Functioning (EF), Pain (PA) and Fatigue (FA). Thresholds for clinical importance were intended to identify symptoms and functional health problems that require a health care professional’s attention.

In this study we have developed clinical thresholds for the physical functioning, emotional functioning, fatigue and pain scales of the EORTC QLQ-C30. By using three anchor items assessing burden, limitations and need for help related with a specific domain, we defined an external criterion for clinical importance. Applying this criterion in our sample, we found prevalence rates for clinically important problems ranging from 22.9 % (pain) to 38.5 % (physical functioning).

To investigate the invariance of diagnostic accuracy across different patient groups we conducted binary logistic regression analyses. These showed that differences across the investigated patient subgroups were minor, indicating invariance of the thresholds across these groups. Based on these findings we do not recommend using different thresholds for different subgroups (e.g. men versus women, patients on- and off-treatment). The invariance of thresholds is an important finding because the use of different thresholds for different patient subgroups would be complex in daily clinical practice and might not always be intuitively understood by health care professionals.

In principle, the anchor items used in our study could be used directly for screening for clinically important problems. However, using the QLQ-C30 and the developed thresholds instead provides additional benefits. The QLQ-C30 provides better measurement precision, in particular with regard to changes over time. It also allows one to relate individual patient scores to norm/reference data available from the literature and to use the collected data for scientific purposes. For these reasons clinicians and researchers may prefer using thresholds for this well-validated questionnaire over only using the anchor items from this study.

The methodological approach most comparable to our study is the analysis by Snyder et al. [8]. They established cut-off scores for the QLQ-C30 using unmet needs defined by dichotomised scores from the Supportive Care Needs Survey (SCNS) as the external criterion. In their study, the diagnostic accuracy of the six investigated QLQ-C30 scales (PF, EF, FA, PA, role functioning and global QOL) was found to be lower than in our study (AUCs were between 0.56 and 0.81). More interestingly, their thresholds were also substantially different from ours, with the differences ranging from 10 to 20 points. In their study, a lower symptom or functional health score was required to qualify as a positive case. This might be related to the use of a different anchor definition that was more sensitive than our TCI criterion. A patient was considered a positive case if s/he had a low, moderate or high need in a certain area, and a negative case if s/he had a satisfied need or if the question was not applicable. In our analysis, a more sensitive criterion like the TMCI would have led to prevalence rates of up to 78 %. In addition to this difference in the definition of the anchor items, the differences observed in optimal thresholds in their study and ours might indicate cross-cultural influences (American versus European cancer patients).

In another study by Snyder and colleagues [18], the QLQ-C30 was also shown to be capable of accurately identifying a patient’s most bothersome issues. For the prediction of these most bothersome issues, sensitivity and specificity for functional health were 0.75 and 0.64, respectively while for symptoms classification accuracy was somewhat better (sensitivity = 0.83; specificity = 0.76). Interestingly, the authors also found that diagnostic performance of change scores was lower than that for absolute scores. This is in line with a recent analysis of Snyder et al. [9] that showed that a change in patients’ QLQ-C30 scores does not relate consistently to a change in patients’ supportive care needs. Such findings support our approach relying on absolute scores rather than on change scores for identification of clinically important problems.

A limitation of our study is that the anchor items and the definitions for (minimal) clinical importance were based on input from an expert panel only. However, the three anchors that we employed in our study reflect characteristics of patients’ symptom burden and functional health limitations that are typically assessed in patient-reported outcome measures, either alone or in combination. Nevertheless, in future research it would be useful to elicit direct patient input on this issues. An additional potential limitation of the study is that the QLQ-C30 pain scale not only assesses severity, but also interference with daily activities. The larger area under the curve observed for the pain scale in comparison to the other three scales that we evaluated probably reflects, at least in part, this overlap of the pain scale with 2 of the 3 anchor items; the other three QLQ-C30 domains assess severity only and had overlap with only one anchor item.

A strength of our study is the large cross-cultural patient sample. In addition, we employed a rather broad definition of what makes a symptom or problem clinically important. We are aware that the set of anchor items is not exhaustive; however, in our study we found the three anchor items to be highly correlated (especially burden and limitations) and it is likely that additional anchor items (e.g. self-rated importance of a symptom) would be highly correlated with limitations, burden and need for help as well. We expect our thresholds to be fairly robust against changes of the criterion as long as correlations with our current anchor items are high.

We believe that the thresholds for clinical importance developed for the four QLQ-C30 scales PF, EF, FA and PA will facilitate interpretation of these QOL scores on an individual- and on a group-level. The thresholds can be used to convert metric QOL scores to symptom prevalence rates in a meaningful way and thus allow comparison with prevalence rates collected with other methods such as clinicians’ ratings. In daily clinical practice such thresholds can guide identification of patients with clinically important problems that should be discussed with the patient. In a next step we will develop thresholds for clinical importance for all QLQ-C30 scales and the corresponding computer-adaptive measures currently being developed by the EORTC Quality of Life Group [21].

AUC, area under curve; EF, emotional functioning; EORTC, European Organisation for Research and Treatment of Cancer; FA, fatigue; HRQOL, health-related quality of life; MIC, minimal important change; MID, minimal important difference; PA, pain; PRO, patient-reported outcome; PF, physical functioning; QOL, quality of life; QLQ-C30, quality of life questionnaire - core 30; ROC, receiver operating characteristic; SD, standard deviation; SNCS, supportive care needs survey; TCI, threshold for clinical importance; TMCI, threshold for minimal clinical importance; UICC, union for international cancer control; UK, United Kingdom