• © 2008 by American Society of Clinical Oncology

Practical Model for Prognostication in Advanced Cancer Patients: Is Less More?

Prognostication of life expectancy is of utmost importance to patients, families, and oncologists, particularly in the advanced cancer setting. Accurate prognostication is essential for oncologists in formulating their recommendations. Patients’ predicted survival can have a profound impact on important decisions, such as palliative care and hospice referral, initiation of specific medications, and avoidance of aggressive therapies. As cancer patients progress over the course of their illness, knowing what to expect can also provide them with a sense of control and facilitate the process of advance care planning.¹ Clinicians have been consistently found to overestimate survival.² Therefore, prognostication tools are needed to enhance the accuracy of survival predictions.

Traditional prognosis for cancer patients is primarily based on tumor-related factors, specifically disease burden and aggressiveness as indicated by various clinical, imaging, laboratory, pathologic, and molecular features. Prognostic models consisting of various combinations of these factors have been validated to provide highly accurate estimates of long-term survival for specific cancer types, allowing oncologists to provide risk-adapted treatment plans and to inform patients of their general prognosis. However, they tend to be less precise when it comes to predicting short-term outcomes in advanced states of disease, limiting their utility in day-to-day clinical decision making for patients with progressive cancer.

To obtain a more accurate estimate of short-term prognosis, newer models have incorporated various patient-related prognostic factors, such as performance status, comorbidities, and physiologic parameters. While a number of prognostic models are now available for advanced cancer patients^3,4 and far advanced cancer patients,^5,6 there are significant barriers preventing their use in routine practice. These include limited applicability due to complexity of the model, restricted generalizability secondary to poor patient selection for the derivation cohort, and overall limited accuracy.

In this issue of Journal of Clinical Oncology, Chow et al⁷ describe the derivation and subsequent validation of a prognostic model for advanced cancer patients referred for palliative radiation therapy. They identified six prognostic factors associated with poor prognosis, and proceeded to evaluate four different prognostic models with various combinations of scoring systems (weighted v nonweighted) and number of variables (3 v 6). The authors favored a three-risk factor model, consisting of nonbreast cancer, metastases other than bone, and Karnofsky performance status ≤ 60 (or Eastern Cooperative Oncology Group performance status ≥ 2), because of its relative simplicity. The median survival rates for patients with zero to one, two, and three risk factors were 55, 19, and 9 weeks in the temporal validation group, and 64, 28, and 10 weeks in the external validation group, respectively.

This proposed model has a number of important advantages, making it an appealing prognostic tool for clinical practice. The authors have validated this prognostic model in two different settings with large patient cohorts, supporting its use in different clinical situations. This model is able to provide a high degree of discrimination among the three groups, with potential practical implications for oncologists. For instance, patients with painful bone metastases and zero or one prognostic factor might benefit from a multifraction course of radiation to minimize the need to retreat for recurrent pain, while patients with two or more risk factors could receive a single fraction. Furthermore, the authors should be commended on their efforts in the development of a highly simplified prognostic model, consisting of only three clinical factors that are reliable and readily available, making it a practical tool in clinical practice.

One of the key issues in the development of prognostic tools relates to the choice of inception cohort, which defines the model's generalizability.⁸ For instance, prognostic factors derived from patients admitted to palliative care units in Canada may not be applicable to those in the United States because of different admission criteria, length of stay, and overall mortality. Chow et al⁷ have wisely chosen to study advanced cancer patients referred for palliative radiation, which represents a relatively homogeneous time point along patients’ disease trajectory. This critical consultation represents a window of opportunity for radiation oncologists to address symptom control issues and, more importantly, advance care planning which requires accurate prognostic information.

The accuracy of the three-risk factor prediction model for individual patients is suboptimal. This is indicated by R² values of 0.24 in the temporal validation cohort and 0.15 in the external validation cohort, suggesting that the three prognostic factors could only partially explain variations in observed survival. Indeed, poor accuracy remains a concern shared by many existing prognostic models. While the limited agreement between observed and predicted survival can be partly attributed to the intrinsic uncertainty with patient related outcomes, study design issues, such as selection of inception cohort and prognostic factors, also likely contribute to this common deficiency.

How can we improve the accuracy of our prognostication? One way is by developing models with more narrowly defined patient subgroups, which will reduce the variability in patient characteristics and thus outcomes. A more specific patient population will also facilitate fine tuning of existing prognostic variables, and the discovery of novel prognostic factors that can provide a higher resolution of accuracy. For instance, there is accumulating evidence that patient-report outcomes carry prognostic significance.

Can this three-risk factor model be applied to advanced cancer patients in other settings, such as patients in medical oncology or palliative care clinics? Potentially yes, although further validation would be required. While performance status alone has been validated in numerous clinical settings, the prognostic utility of tumor type and site of metastases may vary depending on patient cohorts. For patients with far advanced cancer (ie, expected survival < 3 months), tumor histology tend to lose its prognostic significance while other patient-related factors, such as performance status, anorexia-cachexia, delirium, and dyspnea, predominate.⁹

The science of prognostication is rapidly evolving. Future models will need to be sophisticated enough to provide highly accurate prognostic information, yet simple enough to allow application by busy clinicians in everyday practice. Routine adoption of such tools, coupled with outcome measures supporting differential clinical practice based on prognostic classification will allow oncologists to formulate individualized management plans and to optimize patient care. However, even with the most sophisticated model, it is important to recognize that there will always be uncertainty in survival predictions due to the inherent nature of cancer deaths, mediated by acute complications such as infections and thromboembolism. Thus, it is imperative for oncologists not only to refine the science of prognostication, but also to further the art of communication, gently guiding patients and families through times of uncertainty.