Development and Validation of a Model Predicting Short Survival (Death Within 30 Days) After Palliative Radiotherapy

Discussion

Estimating life expectancy in patients with advanced cancer is not trivial. In a study by Hartsell et al., physician-predicted survival was optimistic compared to actual survival, by an average of three months (14). Finding objective criteria which physicians can rely on is, therefore, of critical importance with regard to reducing over/undertreatment in this patient group (15-17). Ultimately, improved models may provide patients with a better understanding of their expected survival and thereby allow them to make informed choices regarding their treatment path at the end of life, whether life-prolonging or symptom-directed palliation (18). In a recent study from Italy, 36% of patients who were admitted to an oncology ward due to acute conditions and died within four weeks were on active treatment (19). In the United States, nearly half of the patients with stage-IV lung or colorectal cancer in the Cancer Care Outcomes Research and Surveillance Consortium had at least one marker of aggressive end-of-life care, including chemotherapy in the last 14 days of life (16%) (20). For the present study, attempts were made to develop a predictive model that might facilitate decision-making specifically for patients considered for PRT. It is clear from previous studies that not all patients benefit from PRT, in part because their survival is too limited to experience symptom relief, which often develops slowly over several weeks (4, 21). Avoiding futile PRT saves resources, both from the patient and provider perspective, since many countries have seriously limited healthcare resources and waiting lists for radiotherapy, avoiding unnecessary PRT might improve overall cancer care. In regions with large distances between radiotherapy centers, patients need to leave their families and usual caregivers, resulting in reduced social support, which is undesirable during the terminal phase. Depending on the healthcare system, the financial burden of accommodation and travelling might be substantial.

Comparable to other studies, we chose to focus on the final 30 days of life, although other definitions of short survival and other measures of futility exist. Prospective data related to PRT of brain metastases from NSCLC in patients with limited survival expectation (median <2 months) suggested no benefit in terms of quality of life or quality-adjusted life years compared to best supportive care (22). A different study from Germany reported that only 17 out of 46 patients were able to complete quality of life questionnaires three months after whole-brain radiotherapy, largely because of short survival (23). In the present study, the median survival of patients who received PRT30 (n=53) was 16 days, with only 10 of these patients (19%) surviving for more than 21 days. Therefore, it appears reasonable to regard PRT30 as a futile approach for the majority of patients. Some patients might possibly benefit from pain relief, improvement of dyspnea or reduced need for medications after PRT, despite short survival. Therefore undertreatment is not desirable either.

Disadvantages of our study include its retrospective design and the fact that patient numbers were limited, especially regarding subgroups, and that most patients were elderly (median age=70 years). Data on potentially predictive cancer-related symptoms, such as cachexia or dyspnea, were not available for analysis. The same is true for post-PRT quality of life and symptom relief data, and cause of death. However, we had access to a large number of variables from a consecutive patient population, representative of everyday PRT practice in most developed countries (including emergency PRT and incomplete PRT courses). Stereotactic radiotherapy was not included in the present series. The majority of PRT courses consisted of hypo-fractionated regimens, mostly 1-15 fractions, with dose/fractionation parameters reflecting a patient's expected prognosis (clinical estimate). We did not use any particular prognostic models or scores when assigning treatment regimens. As evident in Table I, patients who died within one month were typically treated with fewer than 10 fractions and at least 4 Gy per fraction (less time- and resource-consuming schedules). We decided to analyze all PRT courses administered during the time period of this study, including repeat courses given to the same patient because each single-course carries a certain risk of shorter than expected survival. Another approach might have been to include only the last PRT course of each patient, i.e. the one with shortest survival. However, this would have resulted in removal of a considerable number of PRT courses which did not result in early death, and reduced statistical power.

In principle, our results indicate that predictive scores might have some value regarding the end-point of PRT30. However, their accuracy was not fully-satisfactory and readers must also be aware that all scores or RPA-related results are based on 330 PRT courses due to incomplete information in the remaining cases. The main reason for not recommending scores is that withholding PRT in all patients with unfavorable prognosis would prevent more than 30%, who actually live longer, from receiving potentially useful symptom palliation. Our RPA-based decision approach with six predictive parameters, namely lung or bladder cancer, ECOG PS 3-4, low hemoglobin, opioid analgesic use, steroid use and known progressive disease outside the PRT volume, was more accurate and appears clinically-applicable due to its lower risk of withholding PRT from patients with longer survival. However, it is applicable only to patients with primary lung or bladder cancer, which constituted 29% of all PRT applied. Although the RPA model was not 100% accurate, it was valid in the independent data set. The latter included 129 courses of PRT whereof 22 (17%) were administered during the final 30 days of life. Nineteen PRT courses (15%) were given to patients with all adverse features. Out of these 19 patients, 14 died within 30 days (74%). Death within 40 days occurred in 16 patients (84%). By using our RPA model, one would withhold PRT in 19 such cases, i.e. 15% of all PRT related to lung or bladder cancer. This figure corresponds to the majority (19 out of 22 courses, 86%) of PRT30. From the point of optimal resource utilization, it would also be desirable to develop decision tools for all remaining primary cancer types because these account for 71% of all PRT applied. This task requires additional studies in larger databases.

Over the time period of our study (4.5 years including our validation data set) we found no decrease in utilization of PRT30. Overall, 9% of all PRT courses were administered to patients with such limited survival. There was no significant influence of age on PRT30, indicating that elderly patients should not be deprived from access to treatment. Based on our results shown in Table I, PRT is also appropriate if one needs to irradiate multiple target volumes at the same time, or if the patient has been irradiated to the same target volume before (re-irradiation). It is important to compare our results to those of other groups. Guadagnolo et al. have reported on use of radiotherapy in the last 30 days of life in the United States (24). They used a Surveillance, Epidemiology, and End Results (SEER)-Medicare linked database to obtain a large study cohort (202,299 patients ≥65 years of age), albeit with limited number of available baseline parameters. Only patients who died as a result of lung, breast, prostate, colorectal and pancreas cancer (top five cancer causes of death) between January 1, 2000, and December 31, 2007, were included. In other words, their study was different from ours, which did not apply age or primary tumor limits, and examined far more predictive variables. The rate of radiotherapy in the last 30 days of life was almost identical (7.6%, potentially few courses of curative radiotherapy might have been included in this study, as discussed by the authors). Multivariate logistic regression analysis revealed that the likelihood of receiving radiotherapy was significantly greater with the following: earlier year of death, lung cancer, younger age, male sex, married status, Charlson comorbidity index of 0, urban residence, neighborhood income level in the highest quartiles, no receipt of hospice care, southern SEER region, and race. No attempt was made to develop predictive models.

Another analysis of a SEER-Medicare linked database (2000-2007, breast, prostate, colorectal and lung cancer, incident diagnosis of stage IV disease) was reported by Murphy et al. (25). Forty-one percent of all 51,610 patients received PRT (median duration 16 days, dose not available). Twenty-three percent of patients with lung cancer died within two weeks of completing PRT (12% with colorectal or breast, 8% with prostate cancer). Other predictors of early death included increased age, increased comorbidity, and male sex. Kapadia et al. analyzed data from the National Comprehensive Cancer Network NSCLC Outcomes database (1,098 patients who died) (26). They looked at receipt of radiotherapy within 14 days of death. Ten percent of patients had received such treatment, a figure confirming that patients with lung cancer are at higher risk. On multivariate logistic regression analysis (fewer variables than our study), independent predictors of receiving radiotherapy near the end of life included stage IV disease or multi-organ involvement at diagnosis, age <65 years at diagnosis, and treating institution. Gripp et al. found that poor PS, shortness of breath, and brain metastases predicted for survival of less than one month (21). Out of 216 patients with different primary tumors referred for PRT in their study, 33 (15%) died within one month. These studies did not attempt to develop predictive models. Given the large differences in patient eligibility criteria and number of available baseline parameters, the discordant results from all these studies are not surprising. Over aggressive cancer treatment at the end of life may be an indicator of poor-quality care, and causes problems for individual patients and also healthcare systems. It is, therefore, important to develop decision tools that facilitate tailored palliative approaches. Our present efforts might contribute to this aim and we recommend additional external validation of our RPA-based predictive model, as well as prospective evaluation of quality of life and symptom improvement in patients with a poor prognosis receiving PRT in order to shed more light on the important question of whether PRT actually achieves its aim. In the clinic, individual decisions have to be made while we await further results. Those prescribing PRT to patients with a poor prognosis should attempt to use short-course regimens and techniques that do not cause toxicity, which might worsen quality of life.