Introduction
Cancer promotes thromboembolism through inflammation and hypercoagulability, and an ischemic stroke may be the first sign of an occult malignancy. Reference Rioux, Keezer and Gioia1 Early recognition of cancer in stroke survivors represents an opportunity to tailor antithrombotic therapy and offer cancer treatments to improve secondary prevention. Reference Rioux, Keezer and Gioia1 The benefits of routine cancer screening, however, are uncertain, partly because estimates of cancer incidence after stroke are conflicting. Additional research with adjustments for potential confounders is needed to reach valid estimates of post-stroke cancer risk. We conducted a retrospective cohort study to compare the risk of cancer in people who experienced an ischemic stroke to that of the general population, using data from the Canadian Longitudinal Study on Aging (CLSA).
Methods
We used the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement to report our work. Reference von Elm, Altman and Egger2
Data Source
The CLSA is a large, national population-based cohort study on adult aging. It consists of 51,338 Canadian women and men aged 45–85 years at the time of enrolment, intended to be followed every three years for up to 20 years or death in one of two complementary cohorts (tracking and comprehensive). Reference Raina, Wolfson and Kirkland3 The comprehensive cohort includes 30,097 participants randomly drawn (mostly using random digit dialing of landline telephones) from people living within 25–50 km of 11 designated data collection sites located in seven Canadian provinces. Participants were asked to report information relevant to health and aging, such as sociodemographics and prior diagnoses. Exclusion criteria at enrolment included people living in institutions and long-term care facilities, as well as those with physical or cognitive impairments limiting their ability to participate with the study. Recruitment and baseline data collection were completed between 2011 and 2015, while the first follow-up period began in 2015 and ended in 2018. Further details on the study methods are available. Reference Raina, Wolfson and Kirkland4
Data Extraction and Retrospective Cohort
We limited our analyses to data from the comprehensive cohort, as a formal definition of stroke was read to participants during data collection (‘when blood flow to a part of the brain stops’) which allowed for a distinction to be made between ischemic and hemorrhagic stroke, while the tracking cohort did not. In addition to stroke, participants were asked to report a prior diagnosis of ‘ministroke’ or transient ischemic attack (TIA). We used the first two questions of the CLSA’s cerebrovascular event algorithm derived from the validated Questionnaire for Verifying Stroke-Free Status (QVSFS) to ascertain stroke and TIA status in our study. Reference Jones, Williams and Meschia5 We extracted data from the first follow-up on current age, sex at birth, past alcohol use, self-reported medical diagnoses (including cancer and stroke) and age at the time of diagnosis. Smoking and ethnicity were extracted from the baseline data.
We defined index age in the exposed group as the age of either TIA or stroke and defined that of nonexposed as the age of matching. When both TIA and stroke were reported by a participant (n = 128), we only considered the first event. We defined incident cancer as any self-reported diagnosis of a first cancer occurring after the index age, excluding non-melanomatous skin cancers (n = 896) given their benign course. Reference Andersen, Olsen and Olsen6 Time from stroke to cancer diagnosis was a priori defined as the difference in the self-reported age at the time of cancer and stroke diagnoses, to adequately manipulate interval-censored data. Reference Stovring and Kristiansen7 We excluded participants whose age at the time of TIA (n = 84), stroke (n = 21) or cancer (n = 74) was missing. We also excluded individuals reporting cancer before (n = 303) or at the same age (n = 36) as that of index stoke to exclude pre-stroke cancer, and those without at least one year of follow-up after the index event (n = 145).
We built our cohort of exposed and unexposed participants using individual exact matching for age. We randomly sampled each participant from the base without replacement, in chronological order by age at the time of the index event. Reference Heide-Jorgensen, Adelborg, Kahlert, Sorensen and Pedersen8 For each participant with stroke at a given index age, four unexposed participants without a prior cancer or stroke were drawn from the base cohort.
Statistical Analyses
We compared baseline characteristics with Student’s t-test for continuous variables and chi-squared test for categorial variables. We calculated the cumulative incidence of cancer per 1,000 individuals along with 95% confidence intervals (CIs) using the Wilson interval for binomial proportions. We compared the first three years of post-stroke cancer incidence with a theoretical homogeneous distribution using the chi-squared goodness of fit test. We used Cox proportional hazards models to estimate the hazard ratios of new cancer diagnosis with and without a prior stroke. We built three parallel models, each with an alternative dependent variable for time of follow-up (any time, four years, and one year after stroke) as we hypothesized that the risk of cancer would be time-dependent. Reference Erichsen, Svaerke, Sorensen, Sandler and Baron9 We identified potential confounders with causal graphs and built a family of three models per dependent variable adjusted for i) demographics alone (sex, ethnicity), ii) demographics plus lifestyle habits (smoking status, alcohol consumption), and iii) demographics and lifestyle habits plus comorbidities (hypertension, diabetes mellitus). We selected the final model for each dependent variable with the Akaike information criterion. We dichotomized the ethnicity variable in our model (white or other) to respect requirements on minimum outcome events per predictor variable. Reference Vittinghoff and McCulloch10 We included index age in our models to adequately account for matching and verified the proportional hazards assumption with graphs and tests based on the Schoenfeld residuals. Reference Sjölander, Johansson, Lundholm, Altman, Almqvist and Pawitan11,Reference Grambsch Patricia12 We performed sensitivity analyses after excluding people who reported a cognitive decline and those reporting a TIA to reduce misclassification of self-reported diagnoses. Reference Prabhakaran, Silver, Warrior, McClenathan and Lee13 We analyzed our data with R Studio (v.1.2) and defined statistical significance as a p-value <0.05. 14
Results
We identified 920 individuals in the stroke group and 3,680 individuals in the age-matched non-stroke group. The stroke group had a significantly higher proportion of common risk factors for stroke (male sex, smoking, hypertension, diabetes mellitus), as well as more frequent comorbidities (myocardial infarction, chronic obstructive pulmonary disease, chronic kidney disease; Table 1). The index event in the stroke group was either TIA (n = 614; 66.7%), stroke (n = 252; 27.4%), or both (n = 54; 5.9%). The median follow-up after the index stroke event was 10 years in the stroke group (interquartile range [IQR]: 4, 17) and 11 years in the non-stroke group (IQR: 5, 19).
Baseline characteristics of participants with and without stroke
Numbers refer to n (%) unless otherwise specified. Bold characters indicate a p-value <0.05.
Abbreviation: SD = standard deviation.
A total of 105 individuals with stroke (11.4%; 95% CI: 9.5, 13.6) and 418 age-matched non-stroke participants (11.4%; 95% CI: 10.4, 12.4) received a new diagnosis of cancer during follow-up, with a median time to diagnosis of 8 (IQR: 4, 17) and 10 years (IQR: 5, 18), respectively. In the first year of follow-up, 14 individuals in the stroke group (1.5%; 95% CI: 0.9, 2.5) and 26 individuals in the non-stroke group (0.7%; 95% CI: 0.5, 1.0) reported a new cancer. As compared to a theoretical homogeneous distribution of cancer diagnoses after stroke, the observed distribution of cancer in the first three years after stroke was uneven (p-value = 0.030), with a higher incidence of cancer in the first year that declined thereafter (Figure 1). The most frequent primary cancers diagnosed in the first year of follow-up were prostate (n = 8; 57.1% in the stroke group versus n = 8; 30.8% in the non-stroke group) and melanoma (n = 2; 14.3% versus n = 3; 11.5%). Less frequent sites were bladder (n = 1; 7.1% versus n = 1; 3.8%), non-Hodgkin lymphoma (n = 1; 7.1% versus n = 2; 7.7%), kidney (n = 1; 7.1% versus none) and other (n = 1; 7.1% versus n = 2; 7.7%).
Cumulative incidence of cancer per year by index event group. This figure shows the cumulative incidence of cancer (per thousand individuals at baseline) per year after the index event, with 95% confidence intervals.
The hazard of cancer in the first year after stroke was significantly higher as compared to people without stroke after adjusting for sex, ethnicity, alcohol use, smoking, hypertension, and diabetes mellitus, with a hazard ratio of 2.36 (95% CI: 1.21, 4.61; p-value = 0.012). The hazard of cancer was not significantly increased beyond the first year (Table 2). Analyses after excluding people with cognitive impairment (71 with stroke and 310 without stroke) yielded similar results (Table 3). The association was no longer significant in the first year after excluding TIAs (n = 614) and age-matched nonstroke participants (n = 2,456), although the hazard ratio point estimates were similar. Visual inspection and nonsignificant Schoenfeld residual-based tests supported the assumption of proportional hazards in all models.
Cumulative incidence and hazard ratios of cancer diagnosis after stroke
* Matched for age.
† Matched for age and adjusted for sex, ethnicity, alcohol use, smoking, hypertension, and diabetes mellitus. Bold characters indicate a p-value <0.05. Abbreviation: CI, confidence interval.
Sensitivity analyses for the cumulative incidence and hazard ratios of cancer diagnosis after stroke
* Matched for age.
† Matched for age and adjusted for sex, ethnicity, alcohol use, smoking, hypertension, and diabetes mellitus. Bold characters indicate a p-value <0.05. Abbreviation: CI = confidence interval.
Discussion
In this retrospective age-matched analysis of the CLSA cohort, the hazard of newly diagnosed cancer in the first year following an ischemic stroke is about 2.4 times higher as compared to people without stroke after adjusting for sociodemographics and shared risk factors. This one-year period at risk after stroke supports a phenomenon of reverse causation, whereby some cancers were truly occult at the time of stroke and acted as component cause. Neoplasms may evolve from several months to a few years before they reach a clinically overt phase, and the interval of greater risk of cancer diagnosis after stroke fits a plausible preclinical phase. Reference Brenner, Altenhofen, Katalinic, Lansdorp-Vogelaar and Hoffmeister15
We identified two published population-based studies that stratified the comparison of cancer risk by year after stroke. Reference Erichsen, Svaerke, Sorensen, Sandler and Baron9,Reference Lindvig, Moller, Mosbech, Moller and Jensen16 Both found a higher risk of cancer in the first year after stroke that declined afterward. These studies, however, used external comparators (from Danish registries) and did not control for important confounders. One study only assessed colorectal cancer Reference Erichsen, Svaerke, Sorensen, Sandler and Baron9 and the other used data collected before the common use of contemporary diagnostic techniques (1977–1984). Reference Lindvig, Moller, Mosbech, Moller and Jensen16 We used, in contrast, an internal comparator which allowed us to control for common risk factors (i.e. sources of confounding bias) and included all cancer types. A more recent cohort study of young people followed up to nine years after a first-ever stroke reported a higher risk of cancer diagnosis overall, although the authors did not explore the risk by time after stroke. Reference Bergman, Henriksson, Asberg, Farahmand and Terent17
Prostate is the most frequent primary cancer site in both groups of our study in the first year of follow-up. Despite its lower thrombotic potential than other sites such as the pancreas, prostate cancer remains a common cause of cancer-associated thromboses because of its high prevalence in the general population. Reference Abdol Razak, Jones, Bhandari, Berndt and Metharom18 In a large retrospective cohort study using the Surveillance, Epidemiology, and End Results program database, the risk of ischemic stroke was significantly increased by about 60% in the first month following a diagnosis of prostate cancer. Reference Navi, Reiner and Kamel19 In a recent review, genitourinary cancers (including prostate) were the second most commonly reported after stroke (18.3% overall), and the most frequent post-stroke cancer in three studies. Reference Rioux, Touma, Nehme, Gore, Keezer and Gioia20 Up to 50% of strokes associated with prostate cancer do not have a determined etiology, a finding that suggests less common causes of stroke may be involved in these patients. Reference Selvik, Thomassen, Bjerkreim and Naess21 Nonbacterial thrombotic endocarditis, a common but underdiagnosed cause of cancer-associated cryptogenic stroke, also occurs in prostate cancer. Reference Taccone, Jeangette and Blecic22
The strengths of our study include a large, population-based cohort and control for common risk factors in the association of stroke and cancer. Our study, however, has limitations. First, a selection bias likely led to an underestimation of the true association between stroke and cancer. Participants in the CLSA comprehensive cohort needed to be relatively mobile and independent at recruitment, reflected by our high proportion of TIAs (66.7%) as compared to hospital-based cohorts (about 15%). Reference Uemura, Kimura, Sibazaki, Inoue, Iguchi and Yamashita23 People with stroke and occult cancer more often die or have a recurrent stroke as compared to those without cancer, which in combination with independence requirements at inclusion in our study likely led to a depletion of cancer-associated strokes. Reference Taccone, Jeangette and Blecic22,Reference Qureshi, Malik, Saeed, Adil, Rodriguez and Suri24 People with a TIA preceding an ischemic stroke, on the other hand, do not appear to have a different risk of incident cancer diagnosis as compared to those with an ischemic stroke only. Reference Erichsen, Svaerke, Sorensen, Sandler and Baron9,Reference YeŞİLot, EkİZoĞLu and ÇOban25 The exclusion of cancers diagnosed at the same age as stroke and better self-rated general health in the CLSA cohort are additional factors that may explain the lower cancer incidence in the first year after stroke (1.5%) as compared to other prospective studies (up to 5%). Reference Kono, Ohtsuki and Hosomi26 Second, the relative distribution of cancer types observed in our study needs to be interpreted with caution. The retrospective design of the study likely led to a depletion of aggressive malignancies with a higher mortality and may explain the absence of pancreas adenocarcinomas in both study groups. Third, a recent stroke or TIA increases the likelihood of a medical contact that may lead to a diagnosis of cancer unrelated to stroke. This surveillance bias would manifest as an increased proportion of cancers diagnosed after stroke. Participants of the CLSA comprehensive cohort, however, report higher socioeconomic status overall than the general population, which is associated with greater cancer screening utilization in the community and lowers the potential effect of a surveillance bias in our study. Reference Rodriguez and Smith27,Reference Colmers, Majumdar, Yasui, Bowker, Marra and Johnson28 Fourth, diagnoses were all self-reported and therefore prone to errors in classification. A diagnostic accuracy study of the QVSFS, however, found a high specificity (99%) and a moderate sensitivity (79%) for self-reported strokes as compared to history and examination by an experienced neurologist. Reference Jones, Williams and Meschia5 Self-reported TIAs also had a high specificity (97%), but a lower sensitivity (45%). Reference Jones, Williams and Meschia5 When compared to medical records, the combination of self-reported strokes and TIAs has a high specificity (99%) and a moderate sensitivity (78%) for ischemic cerebrovascular events. Reference Okura, Urban, Mahoney, Jacobsen and Rodeheffer29 The effect of this bias on our results was likely small as the hazard ratio point estimates were similar after excluding TIAs, although the few participants left in the analyses yielded non-significant results.
Conclusion
In this retrospective cohort study from the CLSA, we found that the one-year risk of a new cancer diagnosis is about 2.4 times higher in people with an incident ischemic stroke as compared to those without a history of stroke. Additional research is needed to determine whether cancer screening tests after ischemic stroke are warranted, and in which sub-populations these may apply.
Acknowledgment
This research was made possible using the data collected by the Canadian Longitudinal Study on Aging (CLSA). Funding for the Canadian Longitudinal Study on Aging (CLSA) is provided by the Government of Canada through the Canadian Institutes of Health Research (CIHR) under grant reference: LSA 94473 and the Canada Foundation for Innovation. This research has been conducted using the CLSA Baseline Tracking Dataset version 3.4, Comprehensive Baseline Dataset version 4.0, Follow-up 1 Tracking Dataset version 1.0, Comprehensive Follow-up 1 Dataset version 1.0, under Application ID 190243.The CLSA is led by Drs. Parminder Raina, Christina Wolfson, and Susan Kirkland.
Disclosures
BR reports no potential conflicts of interest. MRK reports unrestricted educational grants from UCB and Eisai and research grants from UCB and Eisai. LCG reports speaker fees and advisory board honoraria from Bayer, BMS Pfizer, and Servier and investigator-initiated funding from Servier.
Statement of Authorship
All authors contributed to the design, data preparation, analyses, interpretation, and writing of the manuscript.
Statement of Ethics
Subjects involved in the CLSA study have given their written informed consent, and our institutional Research Ethics Board approved this secondary analysis.
Disclaimer Statement
The opinions expressed in this manuscript are the authors’ own and do not reflect the views of the CLSA.
Data Availability Statement
Data are available from the CLSA (www.clsa-elcv.ca) for researchers who meet the criteria for access to deidentified CLSA data.
Supplementary Material
To view supplementary material for this article, please visit https://doi.org/10.1017/cjn.2021.55.
