INTRODUCTION — Survival estimates are critical factors in clinician and patient decision making in all phases of a serious and/or life-threatening illness. However, prognostic estimates in patients with advanced terminal cancer may have increased importance as they approach the end of life since this is a natural time to formally reevaluate the goals of treatment, with palliative care becoming more prominent and disease-directed therapy less so.
However, despite its importance, prognostication in advanced terminal cancer is imperfect. Clinicians are typically optimistic in their estimates of patient survival, and the prognostic estimates they communicate to patients may be even more optimistic. Some of this misinformation may also be attributed to a lack of communication between oncologists and their patients about estimated life expectancy. In addition, many patients and families may avoid processing and internalizing bad news as an understandable (and potentially adaptive) response to bad news. Thus, effective communication on prognosis in terminal cancer is important but challenging . (See 'Accuracy of survival estimates' below and 'Communication of survival estimates and prognostic awareness' below.)
In an effort to improve these estimates, investigators are integrating previously established prognostic factors into easy-to-use models that clinicians can use in the clinical care of their patients with advanced terminal cancer. The goal of improved prognostication is to provide patients with a better understanding of their expected survival, thereby allowing them to make informed medical and social choices regarding their treatment path at the end of life, whether disease-directed, palliative, or a combination of both [2,3].
Here we discuss general aspects of estimating survival in patients with advanced terminal cancer. Prognostic estimates and the factors that influence outcome in specific cancers are discussed in separate topic reviews that cover individual tumors.
ACCURACY OF SURVIVAL ESTIMATES — Survival estimates that clinicians make, guided only by their intuition and clinical experience, are often incorrect, and the direction of the error is almost always optimistic [4-12]. That is, clinicians tend to believe that their patients have longer to live than they actually do. The disparity in actual versus estimated survival varies in different studies (table 1) [4-8,13-19]. In an illustrative study, 343 physicians referring 468 patients to one of five Chicago-area hospices were asked to provide an estimate of their patient's prognosis ("What is your best estimate of how long this patient has to live?"), and they compared these prognostic estimates with actual survival . The median observed patient survival was 24 days, but the mean ratio of predicted to observed survival was 5.3.
Nevertheless, the clinician's intuitive estimates are of value, particularly when integrated with other methods to estimate the length of remaining life [6,7,20,21]. As an example, in one study, multivariate regression models that included physicians' prognostic estimates were more accurate than models without physician input . Others have shown that clinicians' predictions for prognosis in advanced cancer are more accurate than those of patients or their caregivers . Thus, while it is true that statistical models can be more accurate than human intuition alone [22,24,25], it is also true that physicians provide important information that is not captured in the models alone.
A common problem for clinicians communicating prognosis is that the individual patient may substantially differ in one or several characteristics from the patients included in the studies yielding prognostic information. This is particularly true for very old and very frail patients. In addition to age, other potentially important factors that commonly differ between patients in clinical practice and patients enrolled in studies may include comorbidity burden, presence of cognitive impairment, and presence of geriatric syndromes. This issue is especially important when utilizing prognostic data from treatment trials and applying it to older patients since these studies often focus on younger patients with fewer comorbidities . Thus, published survival estimates from studies should be used only as an informative anchor and starting point for a quantitative estimate of survival. However, clinicians should adjust these published survival estimates if the patient differs substantially from the studied patients in factors that affect survival.
COMMUNICATION OF SURVIVAL ESTIMATES AND PROGNOSTIC AWARENESS — Despite the inability of clinicians to accurately predict survival, most patients expect open and honest communication regarding their prognoses. In a survey of 126 patients recently diagnosed with cancer, 98 percent wanted their doctor to be realistic about their prognosis and provide an opportunity to have their questions answered . The clinician behaviors that were most important to the patient included offering information about current treatments, being knowledgeable about the cancer, and providing reassurance that pain would be controlled. On the other hand, patients with advanced cancer are commonly ambivalent about prognostic information, especially as death seems to be imminent . Patient preferences for disclosure of serious news are discussed in detail elsewhere. (See "Discussing serious news", section on 'Patients’ preferences when receiving serious news'.)
Although medical oncologists report routinely informing their terminally ill patients that they will die, many do not routinely communicate an estimated survival time to their patients . In an analysis of 590 patients with advanced cancer (median survival 5.4 months), 71 percent wanted to be told their life expectancy, yet only 18 percent recalled a prognostic disclosure by their physician . Patients who report recent discussions of prognosis and life expectancy with their oncologists have a better understanding of the terminal nature of their illness .
For those who do provide an estimate of life expectancy, just as there is unconscious optimism in the intuitive estimates that clinicians make about their patients' prognoses, there is also additional, and likely more conscious, optimism in the prognostic estimates that clinicians actually communicate to their patients. This was illustrated in a study in which investigators asked clinicians referring terminally ill cancer patients for hospice care how long they thought their patient had to live . They also asked clinicians what prognosis, if any, they would communicate to the patient if the patient were insistent on receiving a temporally specific estimate. The median survival that clinicians estimated for their patients was 75 days, the median survival that clinicians would communicate to patients was 90 days, and the median observed survival was actually only 24 days.
Finally, even when clinicians believe that they have had completely thorough discussions about prognosis with patients, there is a high frequency of discordance between patients' opinions about their prognosis and that of their oncologist [1,33,34]:
●In a systematic review and meta-analysis of accurate prognostic awareness in advanced/terminal cancer (n = 34 studies), only 49 percent of patients accurately understood their prognosis. Accuracy was highest in Australia (n = 3 studies, 68 percent) and lowest in Italy (n = 6 studies, 30 percent). Accuracy in the United States was 49 percent (n = 7 studies) .
●Similarly, in a cross-sectional analysis of 236 patients with advanced cancer and 38 oncologists from nine different practices, two-thirds of the patients had opinions about prognosis that differed from that of their oncologist, nearly all were more optimistic than the oncologist, and only 1 in 10 discordant patients knew that their opinions differed . Discordance was even more common among non-white patients.
These data suggest that there is a great need for improvement in physician-patient communication in the context of advanced cancer. A number of studies have evaluated interventions to improve prognostic discussion and end of life planning, but while many of these demonstrate efficacy in prompting more frequent prognostic discussion, few have directly measured prognostic awareness [1,36]. A range of patient factors may promote prognostic awareness, including exposure to symptoms, being free of pain and psychological stressors, an active coping style, and family acceptance. Doctor and patient communication styles that allow open, individualized discussion that contributes to mutual understanding as well as a trusting relationship, may also facilitate prognostic awareness .
Additional information on communicating prognosis in palliative care patients (including when to discuss it and how) is presented separately. (See "Communication of prognosis in palliative care".)
IMPROVING PROGNOSTIC ACCURACY — Within clinical oncology, there is a growing literature focused on identifying clinical predictors of survival for advanced cancer patients. These predictors go beyond the basic demographic and tumor characteristics collected in epidemiologic studies of cancer (see cancersurvivalrates.com for calculators based on age, gender, cancer type, and tumor characteristics). Multiple prospective and retrospective cohort studies have identified several clinical predictors of patient survival, most notably performance or functional status and clinical signs/symptoms.
However, an important point is that molecularly targeted therapy and immunotherapy are changing the landscape of treatments for a subset of advanced cancers. As we describe below, with these newer treatments, immunohistochemistry and molecular markers portend substantial variation in response to these treatments and prognosis.
Researchers are seeking to integrate these factors with each other and with new prognostic factors to develop easy-to-use composite measures for the clinical care of patients with advanced terminal cancer. Although a number of tools have been developed, there are few comparative studies. A systematic review concluded that further studies comparing all proved prognostic markers in a single cohort of patients with advanced cancer are needed to determine the optimal prognostic tool .
Predictors of survival
Performance status — Performance status is a measure of a patient's functional capacity . It has consistently been found to predict survival in patients with cancer, and it is used in that capacity as an entry criterion and an adjustment factor for clinical trials of anticancer therapies. A number of metrics have been developed to quantify performance status; among them, the Eastern Cooperative Oncology Group (ECOG) performance status (table 2) and Karnofsky Performance Status (KPS) (table 3) are the most commonly used.
The KPS ranges from values of 100, signifying normal functional status with no complaints nor evidence of disease, to 0, signifying death (table 3). Numerous studies have described an association between survival in patients with cancer and their KPS [4,6,10,15,39-53]. The magnitude of the association is described differently depending upon the statistical methods employed in various trials, but a KPS of less than 50 percent consistently suggests a life expectancy of less than eight weeks for patients enrolled in palliative care programs (table 4) [4,6,15,40,54].
The Palliative Performance Scale (PPS), which includes information about self-care, oral intake, physical activity, disease extent, and level of consciousness, has been found to have similar predictive accuracy for survival as the KPS [51,55-60]. Since the KPS and the PPS appear to be highly correlated, these two scales may be used interchangeably in many populations .
Clinical signs and symptoms — The utility of clinical signs and symptoms as independent prognostic factors was first described in 1966 . In particular, dyspnea is strongly and inversely associated with survival in terminally ill cancer patients [10,15,40,50,63-65]. (See "Assessment and management of dyspnea in palliative care".)
In a qualitative systematic review examining over 100 different variables from 22 studies of patients with advanced cancers, the authors reported that, after performance status, certain discrete signs and symptoms were the next best predictors of patient survival . Dyspnea, dysphagia, weight loss, xerostomia, anorexia, and cognitive impairment had the strongest evidence for independent association with cancer patient survival in these studies. The range of median survivals for the various signs and symptoms reported in univariate analyses from these and other studies is presented in the table (table 4). These findings suggest that for patients with advanced terminal cancer, such as those referred to palliative care programs, the presence or absence of these symptoms may help clinicians estimate patient survival.
Integrated prognostic models — Increasingly, clinical researchers are developing easy-to-use prognostic models that combine elements from established prognostic domains (eg, clinicians' clinical predictions, patient performance statuses, patient-reported signs and symptoms, primary and metastatic sites) to render more accurate survival estimates. To select the most appropriate prognostic tool for a given patient, clinicians need to understand the characteristics of the population in which each tool was developed and validated to determine whether it is applicable (generalizable) to the individual patient.
Broadly, tools have been developed through study of at least two distinct types of patients with advanced cancer: those who have already been referred for purely palliative care and those who may still be receiving, or at least are still candidates for, disease-directed therapy. While models have been derived in these two different patient populations, it is not clear that prognosis differs all that much. In fact, at least two studies demonstrate that patients receiving early palliative care in conjunction with disease-directed treatments have extended survival [66,67]. (See "Benefits, services, and models of subspecialty palliative care", section on 'Rationale for palliative care'.)
A systematic review of cancer presentations with a median survival of six months or less concluded that there was little evidence that disease-directed treatment improved survival in the terminal stages of disease . The authors established a series of clinical factors that were associated with a median survival of six months or less in a variety of tumor types (table 5).
Prognosis in patients receiving palliative care only — For patients choosing to enroll in palliative care, the Palliative Prognostic Index (PPI) provides objective estimates of likely survival (table 6) [64,69]. The PPI, which was developed in advanced cancer patients who had already enrolled in palliative care, can predict three-week survival with a sensitivity and specificity of 83 and 85 percent, respectively, and six-week survival with a sensitivity and specificity of 79 and 77 percent, respectively .
Several other groups have developed similar scoring systems that rely on integration of all or some of the previously described classes of prognostic indicators of patients with advanced cancer who are already receiving palliative care [16,50,70-75]. One of the most widely used is the PPS, which is used in the e-prognosis tool to estimate one-year survival in patients with advanced cancer. The PPS [55,76] may be a better indicator of functional status than other types of performance status scales, including Karnofsky and ECOG, in this population. Comparisons suggest there is considerable overlap .
Other investigators have used information on ECOG performance status (table 2), time from initial diagnosis to diagnosis of terminal disease, serum albumin, lactate dehydrogenase (LDH) concentration, and lymphocyte count from 406 terminally ill cancer patients to construct a nomogram (figure 1) to predict the probability of survival at 15, 30, and 60 days . The predictive model was validated in an independent dataset of 474 terminally ill cancer patients. Prognostic scores derived from these clinical prediction models are all highly predictive for patients with advanced terminal cancer who are receiving supportive care alone, and there is no strong evidence that one model is superior to another .
Another place for clinicians to look for information about survival in advanced cancer is in the survival curves from studies that include patients who are not treated with anticancer therapy. Natural history studies and randomized clinical trials that include a "best supportive care" arm are two such examples. Natural history studies tend to be single-institution case series of untreated patients with mortality follow-up. Median survivals from natural history studies in breast cancer , head and neck cancer , and hepatocellular cancer  and from patients on the "best supportive care" arms of randomized clinical trials in advanced non-small cell lung cancer [83-86], stage IV colorectal cancer , stage IV or recurrent head and neck cancer , unresectable hepatocellular cancer [88,89], stage IV pancreatic cancer , and stage IV gastric cancer  are shown in the table (table 7).
Prognosis in patients continuing to receive anticancer therapy — Survival studies focusing on patients who are still receiving anticancer therapy (ie, chemotherapy and/or radiotherapy) and not yet referred for formal palliative care may provide objective information on the most likely survival experience for patients who are still healthy enough to be continuing to receive anticancer therapy [49,52,92-96]. However, as noted previously, there is little evidence that the survival experiences of patients electing to pursue a purely palliative approach is different compared with patients electing to continue anticancer therapy.
One such study of patients referred for palliative radiotherapy utilized a model composed of only three predictor variables (ie, KPS [>60 versus ≤60], location of the primary cancer [breast versus non-breast], and site of metastatic disease [bone only versus others]) to stratify these radiotherapy patients into three prognostically separate groups with relatively short median survivals (10 to 11 weeks), moderate median survivals (ie, 21 to 29 weeks), and longer median survivals (ie, 53 to 64 weeks) .
Another study used KPS, number of metastatic sites, low serum albumin, and LDH concentration to stratify 177 hospitalized patients with solid tumors into three prognostic profiles with short (≤two months, with no patient alive at four months), intermediate (25 percent alive at four months), and long survival (80 percent alive at four months) .
As noted above, an important concept that is not reflected in these survival estimates is that within a certain cancer stage and type, prognosis can vary dramatically depending on molecular features. As an example, in advanced non-small cell lung cancer, the presence of an epidermal growth factor receptor (EGFR) mutation confers a more favorable prognosis and strongly predicts for sensitivity to an EGFR tyrosine kinase inhibitor (TKI). (See "Personalized, genotype-directed therapy for advanced non-small cell lung cancer".)
While some patients receiving genotype-directed therapy can have exceptional responses, other patients have disappointing responses. The best way to communicate prognosis across this wide spectrum of response in this new era of molecular feature-directed cancer therapy is unclear .
Prognostic consultation from colleagues — Research from a number of groups suggests that prognostic estimates from colleagues may be more accurate than estimates by the treating clinicians. Two studies found that survival predictions averaged across clinicians were more accurate than a prediction from a single physician [98,99]. Another study revealed that physicians who know little about the patient may provide more accurate predictions than physicians with an emotional or other stake in the outcome of a patient's care, and that older physicians (those in the upper quartile of practice experience) rendered the most accurate predictions . The cumulative evidence from this research suggests that disinterested senior physicians may be a good source of prognostic information. More generally, through "curbside" consultations or more formal avenues like tumor boards, clinicians may find colleagues who are able to more accurately determine patient prognoses.
Machine learning algorithms — Most of the existing integrated prognostication tools, which are based on logistic regression analysis, do not accurately identify most patients who will die in the short term (ie, ≤6 months). Such patients might be deemed appropriate for conversations about treatment and end of life preferences if they have not already taken place, and they might also be appropriate candidates for hospice enrollment. (See 'Accuracy of survival estimates' above and "Discussing goals of care", section on 'Timing of the discussion' and "Hospice: Philosophy of care and appropriate utilization in the United States", section on 'Eligibility'.)
The application of machine learning technology may allow more accurate prognostication by modeling both linear and nonlinear interactions among many variables [100-102]. Two different machine learning approaches were compared with conventional logistic regression of data derived from the electronic health record in an analysis of 25,252 patients with cancer who had an outpatient encounter at a large academic cancer center or an affiliated community practice over a five-month period . The collected data included demographic (age and sex), comorbidity, cancer-related (metastatic cancer versus not, solid versus nonsolid tumor), laboratory, and select electrocardiogram variables. Compared with logistic regression modeling, positive predictive values for death at 180 days were higher with both the machine learning approaches (51 and 49 percent versus 45 percent), and the machine learning algorithm was able to separate patients into high and low prognostic groups (180-day mortality 51 versus 3 percent, 500-day mortality 64 versus 8 percent). In a survey of 15 outpatient clinicians who reviewed the results of the machine learning analysis in "real time," the majority of patients identified as high risk were deemed appropriate for a conversation about treatment and end of life preferences.
The authors subsequently prospectively validated the machine learning algorithm in 18 oncology practices within the same large academic health care system . At a prespecified 40 percent risk of mortality threshold to differentiate high- versus low-risk patients, the observed mortality was 45.2 percent (95% CI 41.3-49.1%) in the high-risk group vs 3.1 percent (95% CI 2.9-3.3%) in the low-risk group. However, while the specificity for this threshold was 99 percent, the sensitivity was only 27 percent, suggesting that many patients with a short life expectancy were missed at this threshold for defining high- versus low-risk patients .
Although this machine learning algorithm holds promise, the lack of external validation preclude us from recommending the use of this or any other similar algorithms. This machine learning algorithm was developed and validated entirely using data from the same electronic health record, raising the possibility that the algorithm was overfit to the data from this single health system and may not perform as well on data from other health systems. Because machine learning algorithms often review hundreds of factors, the risks of overfitting to the idiosyncrasies of local data systems is substantial. Before these algorithms can be recommended, they should be validated using electronic health record data from other health systems.
While machine learning approaches such as these appear to slightly improve predictive accuracy by using the wide range of clinical data available in electronic health records, they sacrificing transparency by developing complex, "black box" algorithms. As more clinical data becomes readily extractable from electronic medical records and clinicians become more comfortable relying on "black box" predictions, machine learning models may become more accurate and more widely used.
●Prognostication in advanced terminal cancer is a difficult but critically important task that will become easier as researchers develop better tools for clinical prediction. Ideally, such efforts will correct the pervasive and systematic optimism in the prognoses that clinicians both estimate and communicate to their patients. (See 'Accuracy of survival estimates' above and 'Communication of survival estimates and prognostic awareness' above.)
●In an effort to improve estimates, investigators are integrating previously established prognostic factors into easy-to-use models that clinicians can use in the clinical care of patients with advanced terminal cancer. The goal of improved prognostication is to provide patients with a better understanding of their expected survival and thereby allow them to make informed medical and social choices regarding their treatment path at the end of life, whether life-prolonging or palliative. (See 'Improving prognostic accuracy' above.)
●Integration of clinician estimates with predictors such as performance status and clinical signs of symptoms (eg, in the Palliative Prognostic Index [PPI] (table 6)) appears to be the best current means of predicting survival among patients receiving palliative care alone. Among untreated patients with advanced terminal cancer, the use of disease-specific survival estimates that are derived from studies in which patients received palliative care only (table 7) may improve predictive accuracy. (See 'Prognosis in patients receiving palliative care only' above.)
●Most prognostication metrics are targeted at patients receiving palliative care; metrics that are focused on patients with advanced cancer who are undergoing active cancer therapy are less common, and it is not clear that any specific tool can be recommended. Furthermore, there is little evidence that the survival experiences of patients electing to pursue a purely palliative approach are different compared with patients electing to continue anticancer therapy.
Accurate prognostication is further complicated in advanced cancer by the fact that within a certain cancer stage and type, prognosis can vary dramatically depending on molecular features. While some patients receiving genotype-directed therapy can have exceptional responses, other patients have disappointing responses. The best way to communicate prognosis across this wide spectrum of response in this new era of molecular feature-directed cancer therapy is unclear. (See 'Prognosis in patients continuing to receive anticancer therapy' above.)
●At least one systematic review of cancer presentations with a median survival of six months or less concluded that there was little evidence that treatment improved survival in the terminal stages of disease . The authors established a series of clinical factors that were associated with a median survival of six months or less in a variety of tumor types, regardless of treatment (table 5). (See 'Integrated prognostic models' above.)
●Machine learning tools may improve the accuracy of short-term prognostication over other integrated models that are based on logistic regression analysis, but further research is needed before they can be assimilated into clinical practice. (See 'Machine learning algorithms' above.)
ACKNOWLEDGMENTS — The editorial staff at UpToDate would like to acknowledge Elizabeth B Lamont, MD, and Nicholas A Christakis, MD, who contributed to an earlier version of this topic review.