We thank Professors Bryan Langholz, Sven Ove Samuelsen, and Yingwei Peng for advice on the case-cohort analysis; Karleen Schulze and Zhi Song for programming and statistical support during the study development and data collection periods; and the Ontario Cancer Registry and Cancer Care Ontario for providing us with the Ontario registry data.
See Editorial on pages 3872–4, this issue.
Treatment choice in prostate cancer is influenced by pre-existing comorbid illnesses, but information about their individual prognostic impact is sparse, and only 1 comorbidity index has been developed for this setting. The authors assessed the impact of individual comorbid illnesses on the risk of early, other-cause death in prostate cancer treatment candidates and propose a modification of an existing comorbidity scale.
A population-based case-cohort study included patients diagnosed from 1990 through 1998 in Ontario, Canada who had planned curative radiotherapy or prostatectomy. The subcohort numbered 1643, and the case sample (those dying of other causes within 10 years) numbered 630. Ontario Cancer Registry data were linked to data from medical charts, including: age, comorbidity using the Cumulative Illness Rating Scale for Geriatrics (CIRS-G), stage, prostate-specific antigen, Gleason score, and treatment. Cox proportional hazards regression assessed the age-adjusted association between CIRS-G and other-cause death.
Respiratory and cardiac diseases were the most common comorbidities and most strongly associated with an increased risk of death. Other important comorbidities included vascular disease, renal disease, and diabetes. The modified CIRS-Gpros score yielded a relative risk (RR) of 1.64 (95% confidence interval [CI], 1.52–1.76) for those scoring 1 compared with 0 and RR 1.18 (95% CI, 1.15–1.21) for each increment above 1. Except for those aged >80 years, results were consistent across treatment type and age group.
Early stage prostate cancer can be cured with surgery or radiation therapy. But these treatments are associated with significant morbidity and, depending on the patient's overall health, the cancer may not progress in his lifetime. Comorbid illness burden is a key consideration in the treatment decision, as many early stage prostate cancer patients die of causes other than the cancer,1 and comorbidity predicts their survival.2-7 In decision analyses, comorbidity has been shown to be a key factor in the prostate cancer treatment decision.8, 9
Most of the research about the role of comorbid illnesses on survival after a prostate cancer diagnosis has used summary comorbidity scores. Our understanding of the role of individual comorbid illnesses is limited to a few studies that either were restricted to prostatectomy-treated patients, studied 30-day mortality, had limited comorbidity information, or had limited statistical power.6, 10, 11
Indexes developed in geriatric or other patient populations do not directly apply to this select group, whose comorbid illness status is generally good because of treatment selection factors.10 Stier and colleagues modified the Total Illness Burden Index for use in the clinic.12 It is a patient self-reported instrument that correlates with physical functioning12 and predicts nonprostate cancer mortality.13 Froehner and colleagues proposed a condensed version of the Charlson index14 that excluded comorbid illnesses they subsequently found were prognostic.10 To our knowledge, no other comorbidity index has been derived for use in this common setting. The Cumulative Illness Rating Scale for Geriatrics (CIRS-G), used in this study, was originally designed for clinical assessment,15, 16 but has also been used successfully in chart review studies17 and in the clinic setting.18 A modified version of this index tailored to this patient population could inform medical history taking in the oncology clinic, could be incorporated into a treatment decision aid, and could be useful as a tool in clinical and health services research.
We used a case-cohort design19 to conduct a population-based study of the role of comorbid illness on other-cause death in men who were candidates for curative prostate cancer treatment. This efficient study design, which oversamples events, enhanced our ability to assess the prognostic role of individual illness severity levels across 14 organ systems using CIRS-G. We then used those individual results to propose a modified version for this setting.
MATERIALS AND METHODS
This retrospective case-cohort study was conducted across the province of Ontario, Canada. We randomly sampled a subset of prostate cancer patients who had received or were candidates for curative treatment (the subcohort). Cases were an oversampling of those who had died from causes other than prostate cancer. Chart data abstractions included detailed comorbidity, disease status, and treatment information. The study was approved by the Research Ethics Board at Queen's University, Kingston, Ontario.
The study population is described in Figure 1. The target population was men with prostate cancer who were treatment candidates and diagnosed between 1990 and 1998 (N = 17,934). They were identified using the Cancer Care Database housed at the Division of Cancer Care and Epidemiology in the Queen's Cancer Research Institute. It combines the Ontario Cancer Registry with treatment information from the regional cancer centers, where all radiotherapy in the province is delivered, and hospitalization data, which include information about prostatectomy.20 Treatment candidacy was defined as having an exploratory lymph node dissection within 7 months of diagnosis (abandoned surgical or, sometimes, radiotherapy treatment because of positive nodes), or a prostatectomy within 7 months, or a curative course of radiotherapy within 9 months. The longer radiotherapy timeframe was chosen because of longer prevailing wait times.21 There were 1068 who died of prostate cancer by our cutoff of 31 December 1999 and 1437 who died of other causes (Fig. 1). The cause of death was determined using the Ontario Cancer Registry, which receives death certificate data on the underlying cause from the Ontario Registrar General.
We stratified our sample by Cancer Care Ontario region to facilitate area-level comparisons for other purposes than this current report. The subcohort numbered 1703 (about 10%) and the other-cause death cases (International Classification of Diseases, 9th edition [ICD-9] not equal to 185) numbered 676 (about 40%). The 611 prostate cancer death cases enumerated in Figure 1 were not included in the current study. After chart review, 58 did not meet inclusion criteria, and 40 had key information missing. By design, the separate sampling for the subcohort and cases resulted in 141 patients being included in both groups. The study population therefore consists of 1643 subcohort members and 630 other-cause death cases and a total study population size of 2132.
Some data come from the Cancer Care Database, but most come from retrospective chart review. Seven data abstractors with medical and/or health records training worked from treating institutions across the province. The study coordinator conducted training and data monitoring. Training included standardized abstractions and an instruction manual that was also used for reference in the field. Data monitoring included on-site chart reabstractions with remedial training. We tested this approach in the pilot study, demonstrating substantial inter-rater reliability.22 Weekly data logic and completeness checks were conducted remotely, and corrections were made by referencing the original chart.
Data were primarily collected from the treating hospital and cancer center charts. If the data were incomplete on key elements, further charts were accessed from other hospitals, urologist, and family physician office charts.
Comorbid illness burden was measured using CIRS-G,15, 16 which we chose based on the results of our pilot study22 because 1) it provided the best resolution of the 5 indexes tested; 2) it explained 7% of the variance in survival, which was higher than the Charlson23 and Kaplan-Feinstein24 indexes and close to the Index of Coexistent Disease25 at 9%; and 3) the abstractors found it easier to use, having the clearest instructions. The CIRS-G collects illness information across 14 organ systems using 4 severity categories for each system. Generally, a 0 score denotes no problem, 1 denotes a current mild or past significant problem, 2 denotes moderate disability or morbidity, 3 denotes a severe problem and 4 denotes an extremely severe problem, such as organ failure or severe functional impairment.26 The 14 severity-specific scores are added to produce the CIRS-G score.15, 16, 27 We used the medical history closest to start of treatment to assign CIRS-G. If the treating chart did not contain a timely medical history, we sought information first from treating physician office charts and then from family physicians.
We captured the TNM classification as recorded in the charts along with additional extent of disease information when necessary to convert to the sixth edition.28, 29 We recorded all available pretreatment prostate-specific antigen results (PSAs) and used the value closest to treatment. Biopsy Gleason score was available for 80%. We used tumor grade to assign a Gleason category to 349 of 434 with missing Gleason scores: Gleason category >7 for poorly differentiated, Gleason 6 for moderately well-differentiated, and Gleason category <6 for well-differentiated tumors as recommended by the American Joint Committee on Cancer.29 We combined stage, PSA, and Gleason data based on the D'Amico risk stratification scheme.30
Treatment information was verified during the chart review. Treatment categories include: pelvic lymph node dissection, prostatectomy, external beam radiotherapy, and prostatectomy with adjuvant radiotherapy. No patients received brachytherapy, as it was rarely used in this era in Ontario.
We describe the patients, their disease, and their comorbidities. We report how often the cause of death was an identified comorbid illness at the time of the prostate cancer diagnosis. We assessed the role of CIRS-G comorbidity score on death from other causes using the Cox proportional hazards model controlling for age. The variance estimates were adjusted for the case-cohort sampling based on the method of Therneau and Li,31 and the area-level stratification was accounted for using an approach and SAS (SAS Institute, Cary, NC) macro developed by Langholz and Jiao,32 with further advice from Langholz about incorporating case sampling weights (personal communication, Bryan Langholz, October 2009). We assessed each organ system separately and then modified the CIRS-G score to: 1) include only those systems associated with nonprostate cancer deaths, and 2) combine severity categories that had similar impact. We investigated the model fit of the continuous modified CIRS-G score by graphically comparing the continuous to the categorical effects for each score from 1 through 12. We also tested the stability of the continuous effect over time (proportional hazards assumption) by adding a time-dependent interaction term in the model. All statistical tests were 2-sided, with alpha = .05 significance level. All data processing and analyses were conducted using SAS version 9.1 for Windows.
Table 1 describes the study population. Most were between 60 and 79 years of age, with a mean of 66.7 years (standard deviation [SD] 6.7) in the subcohort and 69.7 years (SD 6.3) in the cases. Thirty percent of the subcohort and 39% of the cases had high-risk disease. Very few aborted treatment after lymph node dissection, and planned adjuvant radiotherapy was uncommon. The radiotherapy to surgery ratio differed between the subcohort and cases. The mean follow-up was 4.4 years (range, 10 days to 10 years).
Table 1. Age, Prostate Cancer Severity, and Treatment Distributions in the Subcohort and Case Groups (%)
RT indicates radiotherapy.
Gleason category was assigned using histologic grade for 349 missing Gleason scores.
We mapped the cases' ICD-9 cause of death codes onto the CIRS-G organ system categories. We combined the heart and vascular disease codes because they are clinically correlated. Table 2 shows the cases' cause of death distribution and the number in each death category with a pre-existing comorbid illness in that category at the time of the prostate cancer diagnosis. The most common death causes were cardiovascular and then respiratory. Eighty percent of those who died of cardiovascular disease had a pre-existing cardiovascular comorbidity, and 92% of those who died of a respiratory cause had a respiratory comorbidity. Other illnesses commonly known at the time of the prostate cancer diagnosis were in the endocrine system (68%) and liver (58%).
Table 2. Cause of Death Distribution by CIRS-G Organ System Categories and the Number With a Comorbidity in That Category at Time of Prostate Cancer Diagnosis
Cause of Death Among the Cases, N=630
Number With a Pre-existing Comorbid Illness in Cause of Death Category
ENT indicates ears, nose, and throat; GI, gastrointestinal; NOS, not otherwise available; NA, not available.
Clinically correlated cardiac and vascular systems were combined.
Other (includes accidents, cancer NOS, and unknown)
The comorbid disease distribution in the subcohort for each CIRS-G disease category is provided in Table 3. The proportion with “no problem” across organ systems varied from a low of 39% for respiratory illnesses to a high of 94% for hematological illnesses, and few patients had a single comorbid disease that scored higher than 2, reflecting the treatment selection process.
Table 3. Comorbidity Distribution in the Subcohort and RR of Other-Cause Death by CIRS-G Organ System Severity Levels
We conducted a separate regression analysis on other-cause death for each organ system controlling for age and for the residual comorbidity score (CIRS-G overall score minus the score for that system). These results are also shown in Table 3. The trends associated with the cardiac and respiratory diseases are most striking, with statistically significant increasing relative risks (RRs) in every severity category. Severe disease was associated with at least a 2-fold risk increase in hematologic diseases (eg, very low hemoglobin and/or white blood cell counts), lower gastrointestinal diseases (eg, diverticulitis, bowel carcinomas), liver diseases (eg, elevated bilirubin, liver function tests >150% of normal), renal diseases (eg, serum creatinine >3.0, pyelonephritis), and neurological diseases (eg, cerebrovascular accident with residual functional hemiparesis or aphasia). Moderately severe disease resulted in a statistically significant increased risk in the vascular diseases, defined as use of daily antihypertensive medication and/or 1 symptom of atherosclerotic disease, and the endocrine diseases, where the most common problem was diabetes requiring insulin or oral agents. As the frequency distributions show, several of these conditions were present in at least 5% of the subcohort, with some exceeding 20%. Several organ systems were not statistically associated with other-cause death.
We developed a modified CIRS-G scoring system using the analyses presented in Table 3. We considered only those organ systems that contained at least 1 statistically significant finding. We combined severity categories when the effects were similar. By using the cardiac and respiratory results as a guide, we assigned a score of 0 to modest, nonsignificant results and to coefficients that were near zero or negative (RR< = 1.0). We assigned a score of 1 to coefficients that were about equal to 0.2 (RR = 1.2), 2 to those around 0.35 (RR = 1.4), 3 to those around 0.5 (RR = 1.65), and 4 to those about 0.75 or more (RR = 2.1). Our CIRS-G (prostate) scoring scheme is provided in Table 4.
Table 4. Severity Scoring System for Revised CIRS-Gpros
Figure 2 shows the modified CIRS-Gpros score distribution. It is skewed, with a range from zero to 15 and a median of 2.0 in the subcohort (interquartile range, 1-4) and 3.0 in the cases (interquartile range, 2-5). Only 17% of the subcohort members and 6.5% of the cases had no comorbid diseases. The CIRS-Gpros score varied across age groups (P < .0001), with a mean of 1.9 (SD = 1.9) in those younger than 60 years of age, 2.8 (SD = 2.2) in the 60- to 69-year-olds, 3.0 (SD = 2.3) in the 70- to 79-year-olds, and 3.4 (SD = 2.3) in the ≥80-year-olds.
Panel A in Figure 3 presents our investigation of the model fit. It shows the coefficients (ln[RR]) for each individual score up to 12 (generated by a categorical indicator variable in a Cox model) plotted against the continuous effect from the model. The distance from a 0 on the CIRS-Gpros scale to a 1 was greater than the distance between subsequent increments on the scale. Adding 2 to CIRS-Gpros scores ≥1 achieves a better fit, as shown in Panel B of Figure 3. Therefore, we recommend making this adjustment when calculating the RR associated with a given score. The resulting formula for calculating the RR for nonzero values of CIRS-Gpros is:
The association between the overall CIRS-Gpros score and other-cause death is presented in Table 5. There was no meaningful difference between the fully adjusted model that included age, T classification, PSA, and Gleason score (CIRS-Gpros RR, 1.17; 95% confidence interval [CI], 1.15-1.20), the crude result (CIRS-Gpros RR, 1.19; 95% CI, 1.16-1.22), and the result presented in Table 5, which controls for age. When we added an interaction term with time, the CIRS-Gpros main effect was reduced from a coefficient of 0.164, which represents the effect at the average follow-up time, to 0.044, which represents the effect near time zero and an interaction effect of 0.028 for each year of follow-up. This CIRS-Gpros–time interaction translates to CIRS-Gpros hazard ratios of 1.08 in the first year, 1.17 in the fourth year, and 1.39 in the 10th year of follow-up. We also explored the consistency of our main finding across age categories. The results were similar for each category aged <80 years (results not shown), but for those aged 80 years and older, there was no indication of increased risk of other-cause death with increasing comorbidity (CIRS-Gpros RR, 0.96; 95% CI, 0.82-1.11). Although the mean score for CIRS-Gpros was slightly higher in the aged ≥80 years group (3.4 vs 1.9-3.0 in the other age categories), their distribution was truncated at a maximum of 7 compared with 13 to 15 in the other groups. Because of the small number in the aged ≥80 years group, the CIRS-Gpros result reported in Table 5 did not change when we excluded them from the main analysis. We also stratified the result by treatment received and found no differences in our overall findings (Table 5).
Table 5. Association Between CIRS-Gpros Score and Other-Cause Death Controlling for Age, Overall and Stratified by Treatment Received
Comorbid illnesses are important to consider when deciding to treat localized prostate cancer, because the disease has a long natural history, thereby increasing the prognostic impact of such illnesses relative to the cancer. It is generally agreed that the benefit of treatment will only be realized if the patient does not die of a comorbid illness within about 10 years.33, 34 As shown in Table 2, many of the illnesses causing death were known at the time of the prostate cancer diagnosis. We identified those comorbid illnesses present in patients who received or were candidates for curative treatment that were prognostic for death from other causes within 10 years. We also modified an existing, well-validated comorbidity index for future use in this setting.
The most outstanding example of the importance of comorbid illnesses is that of respiratory disease (which included having a significant smoking history), where 60% of the study population had an increased risk of other-cause death because of mild (26% increased risk), moderate (55% increase), or severe (64% increase) disease. Also striking was the role of mild through severe cardiac disease, present in 28% and associated with between 26% and 61% increase in risk. Other key effects include: moderate vascular disease, present in 24% and associated with a 19% increased risk; mild renal disease, present in 10% and associated with a 21% increase; and medication-dependent diabetes, present in 7% (which is likely higher today) and associated with a 35% increase. Conversely, we also identified important effects of uncommon comorbid illnesses. Examples include: severe lower gastrointestinal illnesses such as diverticulitis, moderate liver problems such as having elevated bilirubin, moderate renal failure as evidenced by serum creatinine levels, and moderate neurological illnesses such as a previous cerebrovascular accident with residual dysfunction.
In addition to the impact of these single comorbidities, many patients had multiple medical conditions that, in combination, produced an increased risk. In those with a CIRS-Gpros score of 3, 54% had 2 illnesses, and 10% had 3. In those with a score of 5, 52% had 3 illnesses, and 3% had 4. In those with a score of 7, 27% had 4 illnesses, and 6% had 5.
Prognostic studies in curatively treated prostate cancer have either not analyzed the role of separate comorbid illnesses, have focused on 30-day mortality, or have had limited data or statistical power, and only 1 modified an index specifically for this group.14 Fowler and colleagues observed that the risk of death within 10 years for curatively treated patients was 66% in those with severe comorbidities, compared with 9% with none.3 Wilt and colleagues, using Veterans Administration administrative data, found that having diabetes or congestive heart failure was associated with 30-day mortality postprostatectomy,6 and Alibhai and colleagues found associations between 30-day mortality postprostatectomy and cardiovascular disease and stroke.4 Simone and colleagues were unable to demonstrate an increased risk of other-cause deaths within 5 years because of cardiovascular disease, diabetes, or renal disease, although they were able to detect effects from current smoking and the presence of another malignancy.11 Froehner and colleagues assessed the prognostic ability of the individual comorbid illnesses in the Charlson score23 in prostatectomy-treated patients.10 Similar to our findings, they observed an increased risk of nonprostate cancer death in patients with congestive heart failure, peripheral vascular disease, chronic lung disease, and diabetes and saw little impact of comorbidity on death in the oldest age group. Tewari and colleagues used the Charlson score in a study of localized prostate cancer that included a group of patients who had conservative management, reducing the score to 2 categories to simplify lookup tables for 10-year mortality risk.35
The CIRS-G is a comprehensive index that has been shown to have content, criterion, concurrent, and construct validity and reliability,17, 27, 36 and specific to this setting, we observed substantial inter-rater reliability in our pilot study.22 As shown in the categorical results in Figure 3, it provides distinct risk sets for almost every observed CIRS-G score. This index allowed us to separately analyze the role of comorbid disease severity across 14 organ systems. After external validation and assessment of its usability in the clinic, our shorter modified version of the CIRS-G could be used to guide medical history taking in the oncology clinic, assess risk in treatment decision aids, and measure and control for comorbidity in clinical research studies. An updated version of the CIRS-G was published in 2008.27 The main substantive changes were that hypertension now has its own category, and other vascular problems are combined with hematological problems. We think it makes the most sense to use our vascular weights for both the new hypertension and vascular-hematologic schemes, because hematologic problems were rare.
We restricted our study population to patients who were candidates for curative treatment either because they had had an exploratory lymph node assessment or because they had received a prostatectomy or high-dose radiation therapy. Our results, therefore, are meant to refine rather than supplant the clinical knowledge that already constrains this group to those with mild or moderate comorbidities. This same study population definition would be more difficult to implement today, because the increasing use of active surveillance makes it harder to identify the curative treatment subgroup using administrative data.
This study has several design strengths. Our cohort was population-based and therefore not subject to referral biases. We focused on the 10-year post-treatment period, which is the key clinical treatment decision timeframe. We increased our study efficiency by using the case-cohort design, which allowed us to oversample the deaths and avoid costly, statistically unnecessary data collection in those who did not have an event, which would have occurred had we used a cohort design.19, 37 This choice produced precise estimates of the effects of the separate organ-specific comorbid illness severity categories present in as few as 5% of the study population.
There are some issues to consider when interpreting our results. Although we sought records of full medical history and abstracted extra charts to find it, the comorbidity detail varied and, on average, likely represents an underestimation of true comorbid illness status. The survival experience of these men, diagnosed in the 1990s, may be worse than today because of improvements in radiotherapy and surgical techniques and increasing numbers diagnosed with earlier staged disease. Improvements in the control of comorbid illnesses, especially cardiovascular and respiratory disease, may mean that we have overestimated their impact. Our scale's restriction to organ systems containing at least 1 statistically significant result may underestimate comorbid illness burden for patients with a score of 4 in the eyes, ears, nose, and throat system and a score of 3 in the genitourinary system. Our results do not apply to those aged ≥80 years, because we did not observe a comorbidity effect in that group. This may be because of small numbers in that group and/or their limited range of CIRS-Gpros scores. This phenomenon is likely the result of more stringent case selection in this age group.10 Although some investigators have documented very good agreement between cause of death from death certificates and from patient charts,38 others document a misattribution bias of “other cause” for those treated aggressively.39, 40 If some of our cases died of their prostate cancer rather than other causes, it likely led to an overestimation of the role of cardiovascular and/or respiratory disease in particular.38, 41
We provide estimates of the impact of individual comorbid illnesses on the risk of other-cause death in prostate cancer treatment candidates. This information alone, or combined through the use of the CIRS-Gpros, has the potential to refine the clinician's ability to advise patients about treatment. After external validation, the CIRS-Gpros could be used in clinical and health services research studies on this population.
CONFLICT OF INTEREST DISCLOSURES
This work was supported by the Canadian Cancer Society (grant number 012,238) and by the Canada Research Chairs Program (P.A.G.).