Comorbidity is an important variable to consider in the decision-making for early stage prostate cancer, as demonstrated by Groome et al. in this issue.1 Comorbidity has been defined as the presence of 1 or more coexisting illnesses that can have an impact on the diagnosis, treatment, or prognosis of a given condition.2 In the setting of cancer, comorbidity has been shown to be, among other things, a powerful predictor of long-term outcomes, particularly for less lethal malignancies, as shown elegantly by Piccirillo et al across a range of malignancies.3
Thinking about comorbidity is particularly relevant in early stage prostate cancer for several reasons. First, prostate cancer is often a slow-growing tumor, with 10-year disease-specific survival well above 90% in recent years.4 As a consequence, competing risks of dying from comorbid conditions become particularly relevant to consider.5 Second, radical, potentially curative treatment of prostate cancer with either surgery (radical prostatectomy) or radiation therapy (external beam or brachytherapy) is associated with significant long-term complications in a typically minimally symptomatic disease. Compounding this issue are the findings that the benefits of aggressive treatment are modest and confined chiefly to those men with a long life expectancy.6 As such, in recent years, there has been a growing recognition of possible overtreatment of a significant proportion of men with prostate cancer, either because their tumor is indolent and unlikely to progress during their remaining years, or because they are likely to die of other causes because of significant comorbidity.7-9 (This is not to say that comorbidity is irrelevant to short-term decisions; it has been shown to have an impact on short-term morbidity and mortality after both radical prostatectomy10 and radiation therapy.11)
To optimize clinical decisions for patients with early stage prostate cancer, clinicians need better tools to: 1) select who is likely to die from progressive prostate cancer without treatment and 2) predict remaining life expectancy by integrating information about age and comorbidity. Although progress has been made in the first area with tools, such as the Cancer of the Prostate Risk Assessment (CAPRA) score12 and a variety of nomograms for specific treatment situations, much less information is available on integrating comorbidity into validated life-expectancy prediction tools.
In this issue of Cancer, Groome et al report an innovative epidemiological study design, the case-cohort design, to assess the impact of individual comorbidities on the risk of nonprostate cancer deaths in a population-based study from Ontario, Canada. From a larger cohort of 17,934 men treated for cure for prostate cancer and diagnosed between 1990 and 1998, the authors identified 630 men (cases) who died within 10 years from causes other than prostate cancer. A total of 611 men who died of prostate cancer were excluded, and 1643 men who were treated for cure but did not die served as a “subcohort” of controls. Data on comorbidity were obtained from chart abstraction, and outcomes (including cause of death information) were obtained from linked administrative databases.
Before discussing and contextualizing the findings, it is worth spending a moment on the case-cohort design. This epidemiological design has been used for more than 2 decades,13 although it is not commonly used in medicine, partly because of perceived computational complexity.14 Its main advantage is the ability to improve efficiency of standard error estimation (ie, precision) over standard case-control designs while simultaneously reducing the number of subjects required compared with a full-cohort design, particularly when the event (in this case, death from other causes) is relatively uncommon. In other words, for example, to examine the contribution of comorbidity to deaths from causes other than prostate cancer, a much smaller total sample is necessary with this design to capture an adequate number of deaths compared with a conventional-cohort design. If the sample is selected in an unbiased way (eg, random sampling from a population), then the risk estimates are unbiased. However, 1 component of the case-cohort design that does not seem well-described in the literature is whether any biases are introduced by eliminating cases that die from competing causes (eg, in this specific study, all men who died of prostate cancer were excluded). This may become relevant if comorbidity led to premature deaths in these men, particularly if cause of death was wrongly attributed to prostate cancer.
Groome et al selected a comorbidity index—the Cumulative Illness Rating Scale for Geriatrics (CIRS-G)—based on its comprehensiveness and based on their prior experience comparing it with 4 other measures of comorbidity,15 in which it performed moderately well in predicting noncancer deaths among 275 patients with prostate cancer. It was also relatively clear and easy for chart abstractors to use in that study,15 in contrast to earlier comments about the CIRS-G by Singh and O'Brien in their review of comorbidity instruments for use in prostate cancer.16 The CIRS-G collects information on comorbidity across 14 organ systems and rates them across a 4-level severity classification. The Groome study featured a fairly large, well-characterized population of 630 middle-aged men who died of other causes. The patients appeared to be a little different from most modern American series in that almost three-quarters were treated with radiation (primarily external beam) and almost 40% had high-risk disease. The most common comorbidities were cardiovascular and respiratory. There was a surprisingly low proportion of men with musculoskeletal comorbidities (eg, arthritis), in contrast with most prospective studies of older adults, suggesting undercoding of some comorbidities in the charts used by Groome et al. The authors found that cardiovascular and respiratory comorbidities were the most common causes of nonprostate cancer deaths, with greater risks observed with greater severity of comorbidity. However, a close inspection of Groome et al's results suggests several severity ratings in some organ systems had relative risks associated with very large 95% confidence intervals (eg, genitourinary, endocrine/metabolic) likely due to relatively few events, suggesting an even larger cohort may be necessary to adequately determine the impact of relatively rare, but potentially severe, comorbidities.
The authors went on to revise the CIRS-G scoring system (CIRS-Gpros) based on their findings and then reported relative risks associated with specific severity levels of comorbidity.
Is the revised CIRS-Gpros ready for prime time use in prostate cancer clinics? No. There are several important limitations, which the authors acknowledge. First, the cohort is dated, and the evaluation and management of both prostate cancer and various comorbidities have changed in the past decade. In particular, the emergence of active surveillance as a treatment approach17 suggests any validation study would require inclusion of a wider cohort of patients with early stage disease. Equally important, without validation of the CIRS-Gpros in an independent sample, its performance cannot be properly evaluated. Another key point that must be considered is whether the CIRS-G (either in its original or updated formulation) is superior to any other comorbidity measure. As the authors point out, several studies have compared 2 or more comorbidity indices for predicting both short-term and long-term outcomes in prostate cancer. My read of the literature (corroborated by some of my own research findings10, 18) is that no single measure has come out as greatly superior for either short-term or long-term prediction, although data are more sparse for long-term outcome prediction. Even in the authors' original article, it is important to note that the CIRS-G was not the “best” measure as measured by the percentage variance explained in proportional hazards models.15 In Groome et al's current article, the CIRS-Gpros was not compared with any other measure of comorbidity, including the original CIRS-G. Although overall measures of goodness of fit or predictive ability are not well-established for survival models, there are alternative methods to provide some useful comparative information across different comorbidity measures (eg, reporting C statistics using logistic regression models). Even the authors' finding that increasing severity of comorbidity may contribute nonlinearly to an increased risk of death must be interpreted cautiously, as this requires both validation in an independent sample as well as demonstration that the added predictive value of integrating comorbidity information in a nonlinear manner is more useful than simpler linear methods. In addition, there are practical implications for such a tool once validated; it may be far easier for a prostate cancer clinician to establish that someone has coronary artery disease or chronic obstructive pulmonary disease in a busy office setting than to attempt to determine its severity, even with guidelines or instruction manuals. Other limitations of the study, including the retrospective study design and reliance on death certificates for cause of death information, are well-recognized.
In conclusion, the authors should be commended for moving the field forward an important step with the use of the case-cohort design and a large, richly described dataset. Although we are some steps away from a predictive tool that can integrate comorbidity into predicting long-term survival, important insights have been provided. The next major steps are to validate the CIRS-Gpros in an independent, more modern, and representative cohort and then to compare it to other measures of comorbidity.