A prognosis is a physicians prediction about a patients future. This prediction may be divided into two distinct elements: foreseeing and foretelling. Foreseeing is a physicians silent, cognitive estimate about a patients illness. Foretelling is the physicians communication of that prediction to the patient. In this article, we will consider both aspects of prognostication as they relate to physicians care of cancer patients as they near death. We will review the findings on the significant inaccuracy of physicians predictions and then offer hypotheses to explain the sources of this inaccuracy. We will describe techniques that may improve physicians prognostic accuracy. Through this review, we hope to show that part of the challenge of providing humane, compassionate end-of-life care to cancer patients may entail accurately foreseeing and foretelling their prognoses.
Because the success of novel anticancer therapies is measured primarily by their ability to extend life, prognosis is a central element of oncologic research. Technologic advances now allow cancer patients to be scrutinized even to the level of gene expression for factors that may explain a comparatively long or short survival. Typically, researchers create statistical models that integrate such factors to predict outcomes, and published results may assist physicians in making predictions and treatment decisions about their own patients.
For example, in a paper from the National Surgical Adjuvant Breast and Bowel Projects first prevention trial (NSABP P-01), Fisher and colleagues developed a risk model, a variant of the Gail model, that integrates a number of proven breast cancer prognostic factors (age, age at menses, age at first parity, personal history of breast disease and/or biopsy, family history of breast disease, and race) to quantify an individuals lifetime risk of developing the disease.[1,2] They used the model to select individuals at high risk of developing breast cancer, and then randomized those individuals to receive daily tamoxifen(Drug information on tamoxifen) (Nolvadex) or placebo. Models, such as this one, that rely on multivariate regression analysis are found in all aspects of cancer research, including translational and basic science research.[3-5]
Although prognosis is a central element of oncologic research, such formal and explicit prognostication is seldom required in the clinical care of cancer patients. There are at least two situations in the care of advanced cancer patients, however, in which physicians need to formally foresee the prognosis: (1) enrollment into experimental phase I (dose-determining) chemotherapy protocols; and (2) referral to hospice programs. Both settings have discrete eligibility requirements pertaining to survival.
Typically, to be considered for entry into phase I trials, patients must have an estimated survival of at least 2 to 3 months, and for entry into a hospice program, patients must have an estimated survival of at most 6 months. Because of these formal requirements, physicians ability to determine fine gradations in survival among cancer patients in their last 6 months of life may mean the difference between aggressive and palliative care.
How Good Are Physicians at Prognostication?
How good are physicians at determining which patients are in their last 6 months of life? Janisch and colleagues analyzed survival data from 349 advanced cancer patients after enrollment in phase I therapies. They found that the median survival was 6.5 months, well above the requisite 2 months described in their eligibility requirements. Overall, approximately 10% of patients died within 2 months, although very few of those with a Karnofsky performance status greater than 70 died before 2 months. Given the low clinical response rates associated with phase I therapies, it is unlikely that survival was enhanced by the therapies themselves. Therefore, results from this study suggest that physicians who enroll patients in phase I protocols are generally able to predict which patients have more than 2 to 3 months to live. An alternate explanation is that other eligibility requirements, such as performance status and laboratory tests, select patients with more than 2 to 3 months to live, obviating the need for the input of physicians. Since the study was not designed to test the prognostic accuracy of physicians, however, it is difficult to draw strong conclusions about the actual role of physician prognostication.
Within the palliative oncology literature, a few studies were specifically designed to determine physicians accuracy in predicting the survival of cancer patients admitted to hospice programs.[7-12] Investigators in these studies measured physicians prognostic accuracy by comparing patients observed survival to their predicted survival. Results of the studies, summarized in Table 1, show that, in aggregate, physicians overall survival estimates tended to be incorrect by a factor of approximately two, always in the optimistic direction.[7-10,12]
Another method for measuring physicians prognostic accuracy is to determine the percentage of patients dying within a calculated interval surrounding their predicted date of death. For example, Parkes identifies extreme errors in prediction by noting that pessimistic errors occur when patients live at least twice as long as their predicted survival and optimistic errors occur when patients live less than half as long as their predicted survival. According to this system, physicians whose patients do not fall into either error category have made correct prognoses (although, admittedly, this is a generous definition of correct). Table 1 also contains a summary of the results of studies using this method of measuring physicians prognostic accuracy.[7,9,10,12] Even with this generous definition of correct, physicians who predict the survival of hospice patients are correct only half of the time. Furthermore, the results show that the direction of these extreme errors is predominantly positive.
Studies of physicians abilities to predict cancer patients survival are not limited to patients in palliative care settings. Physicians prognostic accuracy also has been evaluated with greater mathematical rigor in ambulatory patients undergoing anticancer therapy. Mackillop and Quirt measured oncologists prognostic accuracy by asking them to first predict ambulatory cancer patients likelihood of cure and then to estimate the duration of survival for patients whose likelihood of cure was zero. At the 5-year point, patients who were alive and disease-free were termed cured; the dates of death of the incurable patients also were determined. Although oncologists were quite accurate in predicting cure, they had difficulty in predicting the length of survival of incurable patients. They predicted survival correctly for only one-third of patients, with the errors divided almost equally between optimistic and pessimistic.
In summary, physicians asked to foresee gradations of survival in advanced cancer patients enrolling in certain therapies (either aggressive or palliative) are able to do so accurately much less than half the time, and, when in error, they tend to overestimate survival. Although clinicians appear to be adept at foreseeing the likelihood of cure in cancer patients, they are not skilled at foreseeing the length of survival in incurable patients.
Two factors may hinder physicians in their attempts to accurately predict survival of advanced cancer patients: the method of prediction used and forecaster bias.
Method of Prediction
There are two general methods of prediction: actuarial prediction and clinical prediction. With the actuarial method, a prediction is made using empiric data contained in life tables.
For example, an oncologist might consult Surveillance, Epidemiology, and End Results (SEER) tables of patients with non-Hodgkins lymphoma (NHL) to predict a given patients 5-year survival. Alternatively, the oncologist might use the International Prognostic Index to determine the likelihood of 5-year survival of a 45-year-old patient with an aggressive stage IV NHL and an elevated lactate dehydrogenase (LDH) level. Models of greater complexity may provide the physician with greater prognostic precision.
In the clinical method, a prediction arises out of human intuition alone, without the benefit of explicit precedent data from similar patients or optimal weighting of patient and disease variables through mathematical formulas.
Although the actuarial method has been shown in many disciplines, including medicine, to be superior to the purely clinical method,[15,16] few actuarial models are designed explicitly to aid physicians in predicting survival of terminal cancer patients. However, several studies have correlated performance status[7,10,17] and symptoms (eg, dysphagia, dyspnea)[17-19] with survival of terminal cancer patients.
A distinct reason that oncologists may be inaccurate in their predictions about the survival of terminal cancer patients may relate to their own biases. Within the literature on prognostication, certain forecaster biases are well-described impediments to accurate prediction. For oncologists, optimistic bias may be the most germane type of forecaster bias. Optimistic bias about personal risk occurs when a person believes that he or she is less likely than others to experience an adverse outcome.
Optimistic bias is pervasive and well studied. A classic example is the uniform optimism held by cigarette smokers about their health.[22-24] In study after study, cigarette smokers rate their personal risk of developing a smoking-related illness far lower than the average smoker, even if they can accurately forecast the risk of smokers in general.
In his review of optimistic bias about personal risk, Weinstein postulates three reasons for such bias that may be applied to physicians caring for patients. First, by employing optimistic bias, physicians invoke denial to shield themselves from a painful reality, perhaps, in this case, the imminence of a patients death. Second, they may think that they are better than their peers (ie, that they take better care of their patients), and, therefore, may believe that their patients will live longer than a survival curve would suggest. Third, optimistic bias may occur because of simple cognitive errors, eg, that a terminal cancer patients performance status of 90 simply means that they cannot die of their cancer in the next 3 months