A prospective study examining cachexia predictors in patients with incurable cancer

Between April 2011 and October 2013, 1739 patients from 30 centres across Europe (27), Canada (2) and Australia (1) were included in the European Palliative Care Cancer Symptom study (EPCCS). The participating centres are presented in the main publication originating from this study [13]. EPCCS was a prospective observational study conducted by the European Palliative Care Research Centre (PRC) (https://​oslo-universitetssyke​hus.​no/​avdelinger/​kreftklinikken/​avdeling-for-kreftbehandling/​prc) and the European Association for Palliative Care (EAPC) Research Network (https://​www.​eapcnet.​eu/​research/​research-network). The overall aim of the EPCCS study was to improve the understanding of symptom development, and how these symptoms may best be assessed and classified in order to improve symptom management. Eligible patients met the following key inclusion criteria: ≥18 years of age; with incurable cancer and attending a palliative care program. Data pertaining to cancer cachexia were retrieved from this dataset and assessed as part of the present study.

Patients were assessed at baseline and then approximately every 4 weeks for at least three follow-up visits or until death. The following information was collected: Demographical data (age, gender, geographical region and treatment setting [inpatient, outpatient, home care]), disease specific data (cancer type and stage [localized vs. metastatic]), height, current body weight and patient reported weight loss in the 6 months prior to inclusion. Weight loss at subsequent visits was calculated by adding measured weight change to baseline reported weight loss. Cachexia was diagnosed at first occurrence of either a) weight loss > 5% since 6 months prior to inclusion or b) weight loss > 2% since 6 months prior to inclusion if current body mass index (BMI) < 20 kg/m². All patients were assessed for cachexia at baseline. If anyone had insufficient data to be assessed for cachexia at baseline, data from the first follow up visit was used as baseline registrations if available. Only patients without cachexia at baseline were included in the analysis.

Performance status was assessed according to Karnofsky Performance Status (KPS) (0–100). Comorbidities in terms of heart disease, renal disease, chronic obstructive pulmonary disease (COPD) or arthritis were registered.

The dataset was not sufficiently large to assess the risk of cachexia development of each individual cancer type. Thus, cancer type was grouped a priori into one of three categories: Low risk - Breast cancer and haematological cancers (lymphoma, leukemia and myelomatosis); High risk – pancreatic and gastric cancer; Neutral risk - all other cancers. This was based on previous literature on cancer type and association with cachexia [14, 15].

The following patient reported outcome measurements (PROMs) were registered: Food intake was assessed as “less than usual”, “more than usual” or “unchanged” according to the abridged Patient-Generated Subjective Global Assessment (aPG-SGA) [16]. Physical and emotional functioning (0 [worst] - 100 [best]), anorexia and fatigue (0[best] – 100 [worst]) were assessed according to the European Organisation for Research and Treatment of Cancer Quality of Life Questionnaire (EORTC-QLQ) C15 PAL [17]. Both instruments are widely used, and well validated in the target population.

Patients who developed cachexia during follow up were identified and the remaining patients were censored at the time of their last body weight registration. Time to cachexia development or censoring was calculated. Univariable Cox Proportional Hazards method was used to estimate hazard rate ratios (HR) for cachexia development with the following predictors assessed at baseline: Age, gender, cancer type (low risk, neutral risk or high risk), cancer stage, comorbidity, weight loss, BMI, performance status, physical functioning, emotional functioning, fatigue, food intake and appetite. All predictors with p-value < 0.20 were included in a multivariable Cox model. Multicollinearity among the candidate predictors was checked by estimating Spearman’s correlation coefficients. The least significant predictors were dropped from the multivariable model one by one in a manual backwards selection, until only significant predictors remained. All possible interactions among the remaining predictors were examined and added to the model if significant. Due to known association between COPD and lung cancer [18], a sensitivity analysis specifically adjusting for lung cancer alongside the a priori cancer type categorization (low risk, neutral risk, high risk) was performed.

In order to construct a model that uses optimal combinations and cut--offs of the identified risk factors to predict cachexia development, a classification and regression tree (CART) analysis for failure time data was used. CART is a non-parametric data-mining procedure which examines all possible cut-offs of every variable to create separate subgroups of significantly different risks. It repeatedly splits the population based on the variable and cut-off that most optimally stratifies risk of cachexia development in the current group. It stops when significant divisions no longer can be performed. The final subgroups were compared, and adjacent subgroups with similar risk of cachexia development were merged to create levels of increasing risk of cachexia development. Kaplan-Meier curves were plotted for each risk-level, and both log-rank test for differences in cachexia development probabilities among risk levels, and test for trend in cachexia development probabilities (i.e. cachexia development probability of Risk-level 1 ≤ Risk-level 2 ≤ … ≤ Risk-level 5) were performed. To assess accuracy of this model, a calibration plot was created by plotting predicted vs. observed risk at 3 months, and Harrell’s c-statistic was estimated to assess discriminatory capacity. In addition, receiver operating characteristic (ROC) curve analysis at 3 months in patients alive and still on study was performed. For the different possible cut-offs of a diagnostic test (all possible cut-offs for the “prediction function” from the model in this case), the ROC curve is a plot of the true positive rate (sensitivity, i.e. the proportion of patients correctly classified by the model among those actually developing cachexia) against the false positive rate (1- specificity, i.e. the proportion of patients wrongly classified as “developing cachexia” by the model, among those actually not developing cachexia). This was done to evaluate if a cut-off with both high sensitivity and high specificity for the prediction of cachexia development could be identified. A two-sided p-value of < 0.05 was considered significant in all analyses, unless stated otherwise. Stata version 13.1 (College Station, Texas, USA) was used for statistical analyses, and the Stata-module Cart [19] was used for the CART-analysis.

The present study demonstrates that information on cancer type, appetite loss and COPD improves accuracy of cachexia prediction when added to the established cachexia classifier, weight loss. This is especially true in patients with no or minimal weight loss (< 3%), whereas in patients with weight loss between 3 and 5%, development of cachexia is imminent, regardless of other factors. Based on these predictors, patients can be stratified into five different risk-levels of cachexia development. Cachexia development is not likely if in Risk-level 1, and conversely, for patients in Risk-level 3 or greater the risk of cachexia development is high. As such, the risk-levels enable the clinician to select which patients must be followed more closely with respect to cachexia development and ensure early adequate therapeutic intervention. To the researcher, this could improve patient selection in intervention trials aiming at preventing cachexia, by including only patients at risk of developing cachexia.

No single cut-off in this five-level risk ladder has both high sensitivity and high specificity of predicting cachexia. Thus, no single criterion was identified that accurately diagnosed pre-cachexia. Future research should attempt to improve prediction of cachexia development, and thereby improve the diagnosis of pre-cachexia. A likely path towards this aim is to examine the role of inflammatory markers in predicting cancer cachexia. Inflammation is a central part of cachexia pathophysiology and considered a driver of cachexia development [2], and it is likely that markers of systemic inflammation would improve accuracy of cancer cachexia prediction. Thus, the addition of inflammatory markers to the predictors identified in the present study is a necessary next step towards diagnosing pre-cachexia.

The strength of this study is that it examines factors related to cachexia development in a large longitudinal cohort of patients, and thus enables the identification of factors that predict cachexia development and their relative importance. As is common in many studies in palliative care, the number of missing follow-up observations was relatively high. It is likely that a worsening in physical condition is among the reasons for patients dropping out, and this would decrease statistical power and may introduce a bias. To mitigate this effect, Cox proportional hazards method was used to let each patient contribute with his or her time on the study. Furthermore, to increase power of statistical analysis, patients with insufficient data at baseline, but with sufficient data at first follow-up visit were included with the latest visit as baseline. This was considered appropriate due to the open study design, which allowed inclusion of patients at any time point in their disease trajectory. The CART method is a data mining procedure that is simple to understand and gives an intuitive result. As the calibration plot (Fig. 4) shows, the resulting risk-level model fit the observed risk very well. This is expected when evaluating the model on data from which the model was developed, and the CART methodology may be criticised for creating models that are overfitted to the data, and thus reduce the external validity. By only including significant factors from the Cox model, the risk of overfitting is reduced, and, in addition, the resulting CART model seems clinically plausible. No objective measurements of body composition were available when assessing cachexia. Effect on weight change by accumulation of third space fluids or shifts between fat and muscle mass could therefore not be assessed. However, the adapted definition used in this study has previously been validated [10], and it could be argued that this definition is of greater clinical practical value as objective measures of body composition not always are available in the clinical setting. As mentioned above, no markers of systemic inflammation were assessed as possible predictors of cachexia development. Although the EPCCS study allowed for registration of incidental C-reactive protein measurements performed within 3 days before inclusion, too few observations were available to enable statistical inferences.