Clinical Gastroenterology and Hepatology 2015;13:148–154
A Rule for Determining Risk of Colorectal Cancer in Patients With Inflammatory Bowel Disease Maurice Lutgens,*,‡ Séverine Vermeire,§ Martijn Van Oijen,*,‡,k Frank Vleggaar,* Peter Siersema,* Gert van Assche,‡ Paul Rutgeerts,‡ and Bas Oldenburg,* on behalf of the Dutch Initiative on Crohn and Colitis *Department of Gastroenterology and Hepatology, University Medical Center Utrecht, Heidelberglaan, The Netherlands; ‡ Veterans Affairs Center for Outcomes Research and Education, and kDivision of Digestive Diseases, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, California; and §Department of Gastroenterology, University Hospital Gasthuisberg, Leuven, Belgium BACKGROUND & AIMS:
Surveillance guidelines for inflammatory bowel disease–associated colorectal cancer (IBD-CRC) are based on findings from retrospective studies. We aimed to create and validate a prediction rule to assist clinicians in identifying patients with IBD who are at low and high risk for CRC.
METHODS:
We performed a retrospective case-control study of 2 cohorts of patients from tertiary care centers (the University Hospital of Leuven, Belgium, and 7 University Medical Centers in The Netherlands). Multivariate Cox regression was used to select independent risk factors for CRC in the Leuven cohort. Based on their regression coefficients (b), we created a rule to predict risk for CRC. In validation studies, the predictive strength was tested by C-statistic analysis and then validated externally in the Dutch cohort.
RESULTS:
In total, we identified 50 patients with IBD-CRC (cases) and 136 patients with IBD without CRC (controls) in Leuven, and 138 cases and 206 controls in the Dutch cohort. From the Leuven cohort we created the CRC risk prediction rule based on 4 risk factors: IBD-type ulcerative colitis (b [ 1.2), primary sclerosing cholangitis (b [ 1.1), extent of colonic disease ‡50% (b [ 1.1), and postinflammatory polyps (b [ 0.8). The prediction rule consisted of a total score for each individual patient calculated from the presence or absence of these 4 risk factors. For example, a score of 13 represented patients who had extensive Crohn’s disease without PSC or postinflammatory polyps, who had a 15% likelihood of CRC in the Leuven cohort and a 24% likelihood of CRC in the Dutch cohort. Scores of 0, 13, 23, 27, and 37 represented patients with Crohn’s disease, and scores 15, 25, 28, 38, 42, and 52 represented patients with ulcerative colitis. The total score per patient had a C-statistic of 0.75. In the Dutch cohort this score had a C-statistic of 0.67.
CONCLUSIONS:
Ulcerative colitis, primary sclerosing cholangitis, disease extent ‡50%, and postinflammatory polyps were found to determine risk for CRC in patients with IBD. A surveillance guideline that incorporates the relative weights of these risk profiles would identify patients at risk for CRC more accurately than algorithms in current guidelines.
Keywords: Colon Cancer; Inflammation; Screening; UC.
he risk of colorectal cancer (CRC) in inflammatory bowel disease (IBD) colitis patients is known to be increased,1 although not as much as previously reported.2 It is clear that the majority of IBD patients will never develop this complication. The challenge lies in identifying patients at particularly high risk by using established and reliable risk factors. Generally accepted factors are a previous diagnosis of colonic dysplasia,3 disease duration,2,4 disease extent,4 and primary sclerosing cholangitis (PSC).5,6 Endoscopic features also can assist in identifying high-risk patients. Both the presence
T
of postinflammatory polyps7,8 and colonic strictures7 have been shown to be associated with an increased cancer risk in ulcerative colitis. Conversely, a normal endoscopic appearance reduces the risk to the same level as Abbreviations used in this paper: BSG, British Society of Gastroenterology; CRC, colorectal cancer; DALM, dysplasia-associated lesion or mass; IBD, inflammatory bowel disease; PSC, primary sclerosing cholangitis; TNF, tumor necrosis factor. © 2015 by the AGA Institute 1542-3565/$36.00 http://dx.doi.org/10.1016/j.cgh.2014.06.032
January 2015
the background non-IBD population.7 The recently updated British Society of Gastroenterology (BSG) guideline9 also includes family history of CRC as a risk factor based on the study by Askling et al.10 Despite the confirmed association of these predictive and protective factors with regard to CRC development in IBD, their combined value in predicting which patients are at a low or high risk of CRC has never been validated. Therefore, current recommendations on risk groups and intervals in the BSG and American Gastroenterology Association guidelines9,11 are based on expert opinion. The aim of the present study was to discriminate reliably between low and high risk of IBD-associated CRC based on combined risk factors. To achieve this we created an internally and externally validated, easy-to-use prediction rule for IBD-associated CRC.
Methods Study Design We performed a retrospective case-control study to identify predictive and protective factors for IBDassociated CRC. To validate findings in an external cohort we collected data from 2 separate cohorts, 1 from the University Hospital Leuven in Belgium and 1 from university hospitals in The Netherlands.12 The Leuven cohort was used to build the prediction rule and the Dutch cohort served as an external validation.
Patient Selection: Leuven Cohort We used an International Classification of Diseases, 9th revision, coding search for the diagnoses of IBD and CRC at the University Hospital Gasthuisberg, Leuven, which is a tertiary referral center in Flanders, North Belgium. Search results were available for the period from September 1999 to August 2009. The search yielded 99 results. After an initial check, we identified and excluded 3 patients with isolated ileal Crohn’s disease, 21 patients with unconfirmed IBD, 7 patients with synchronous diagnoses of IBD and CRC, 10 patients with low-grade dysplasia but no CRC, 1 patient with high-grade dysplasia but no CRC, and 7 patients with cancers other than colorectal adenocarcinoma. To expand our cohort we cross-referenced the local electronic patient database with all local pathology reports from September 1990 through June 2011. This yielded 5 additional CRC cases between 1990 and 1999. Controls in Leuven were selected from the Leuven IBD biobank by generating a randomly ordered patient list. Controls then were selected consecutively from the top of the list. Crohn’s patients with only ileal involvement were excluded (n ¼ 12). In total, we collected data from 50 cases and 136 unmatched controls. We used unmatched controls to include as many modifiers of CRC risk as possible.
Predicting CRC in IBD
149
Patient Selection: Dutch Cohort The cohort from The Netherlands included all IBDassociated CRC cases from 1990 to 2006 in tertiary referral centers. The cohort of Dutch patients with IBDassociated CRC has been reported previously.12,13 These patients were selected using the nationwide pathology automated archive (known as PALGA).14 Search terms for colitis and carcinoma with multiple synonyms were used to identify patients with IBD and CRC. Results were screened manually for confirmed IBD and CRC. By using this method we identified 149 IBD-associated CRCs in 7 Dutch university medical hospitals. Eleven patients were excluded because of synchronous diagnoses of IBD and CRC. We selected controls by using identical search terms, but this time we excluded neoplasia. A random number generator was used to select controls in a 1:2 ratio to cases. After manual screening, 94 selected controls were excluded because of isolated ileal Crohn’s disease or an unconfirmed diagnosis of IBD. In total, we collected data from 138 cases and 206 unmatched controls.
Data Collection The starting point of data collection and follow-up evaluation was the date of symptom onset that could be attributed to IBD. This was defined by a persistent change of bowel habits and/or bloody diarrhea and/or continuous abdominal pain followed by a diagnosis of IBD. If no clear onset of symptoms was recorded, the date of IBD diagnosis was used. The end of follow-up evaluation for cases was the date of CRC diagnosis. The end of follow-up evaluation for controls could be any of the following: (1) end of study date, which was October 15, 2011, for the Leuven data, and July 1, 2006, for the Dutch data; (2) date of death by any cause; (3) loss to follow-up evaluation defined by the date of the last known visit to the outpatient clinic; or (4) the date of total colectomy. The following variables were collected for both cohorts: sex; IBD type, defined as ulcerative colitis or Crohn’s colitis (indeterminate colitis cases were analyzed with the ulcerative colitis group); smoking, defined as positive for active smoking at the end of the follow-up evaluation or a smoking history; family history of CRC, defined as any first-degree relative having a diagnosis of CRC; limited disease, defined as microscopic disease extent of less than 50%, and extensive disease, defined as microscopic disease extent equal to 50% or greater of the colonic surface; concurrent PSC; postinflammatory polyps; dysplasia-associated lesion or mass (DALM); adenoma-like lesion; flat low-grade dysplasia; flat high-grade dysplasia; colonic stenosis; any mesalamine use; any thiopurine use; any anti–tumor necrosis factor (TNF)-a use; any methotrexate use; colonoscopic surveillance, defined as a dichotomous variable (as yes when random biopsy specimens were taken every 10 cm or chromoendoscopy was performed;
150
Lutgens et al
as no if not); pretumor colonic resection; and age at IBD diagnosis (divided as 36 y). These age cut-off values were based on age distribution within the Leuven cohort.
Statistical Analysis Overall risk factor analysis in the Leuven cohort. Before establishing the prediction score, we first identified risk and protective factors independently associated with CRC. A multivariate Cox regression analysis was used to address different follow-up times between individual patients. This was followed by stepwise conditional elimination using the following categoric variables: sex, IBD type, smoking behavior, microscopic extent more or less than 50% of the colonic surface, PSC, postinflammatory polyps, DALM, adenoma-like mass, flat low-grade dysplasia, flat high-grade dysplasia, colonic stenosis, any mesalamine use, any thiopurine use, any anti–TNF-a use, any methotrexate use, colonoscopic surveillance, pretumor colonic resection, age younger than 19 years, and age older than 37 years. Entering only these 2 age categories and not the middle category of age 19 to 36 years enabled a comparison between the hazard ratio of the younger and older groups with the middle group. The results of this analysis provided an overview from which we could select factors to use in the prediction score.
Prediction Rule The main goal was to create a prediction rule with a minimal number of risk factors. Hence, we only selected independent factors associated with CRC in the Leuven cohort that would be easy to use in clinical practice. We excluded possible protective factors such as colonic resection, colonoscopic surveillance itself, and medication use. We additionally excluded flat low-grade dysplasia, flat high-grade dysplasia and DALM, because these variables are outcomes of surveillance, regardless of their association with CRC. We used multivariate Cox regression analysis with stepwise elimination to identify the remaining factors that had the strongest association with CRC in the Leuven cohort. Only factors with a final P value of less than .05 were included in the prediction rule. The selected risk factors were assigned weights that were derived from their regression coefficients (b). After division by the smallest b, all bs were converted to integer scores, for example, a b of 2.2 corresponded to 2 points and a b of 14.3 corresponded to 14 points. The prediction rule consisted of the total additive score per patient based on the presence or absence of these risk factors. A higher total score per patient indicated a higher risk of developing CRC. To test its discriminative power, we calculated a C-statistic for the total score of each individual patient.15,16 To this end, we used a Survival Analysis SAS Macro that is publicly available from the
Clinical Gastroenterology and Hepatology Vol. 13, No. 1
Mayo Clinic (http://www.mayo.edu/research/departm ents-divisions/department-health-sciences-research/divi sion-biomedical-statistics-informatics/software/locallywritten-sas-macros). We report C-statistic values based on calculations with ties. Calibration is the degree of correspondence between the estimated probability produced by the score and the actual observed outcomes. We assessed the calibration performance by comparing the predicted risk for CRC with the observed number of CRCs in the Leuven cohort (internal validation) and the Dutch cohort (external validation). SPSS version 20 (Armonk, NY) and SAS version 9 (Cary, NC) software were used for all analyses.
Results Patient Characteristics Table 1 shows the overall distribution of variables in the Leuven and the Dutch cohorts. In both cohorts there were more ulcerative colitis and male patients with CRC. Disease duration was longer in the CRC groups. Active smoking or a smoking history was more prevalent in the control groups. A family history of CRC was slightly more prevalent in the Leuven control group and the Dutch CRC group. Disease extent, PSC, and postinflammatory polyp ratios were similar in the Leuven and Dutch cohorts with more extensive disease, PSC, and postinflammatory polyps than in the CRC groups. DALMs were seen only in the CRC group. The Dutch control group had no dysplasia because those patients specifically were excluded when the data for this cohort were collected. More CRC cases than controls had a colonic stenosis, although a colectomy was more common in controls. Especially in the Leuven cohort, more controls were on thiopurines and methotrexate medication and were treated with anti–TNF-a. Only the Dutch cohort showed a difference in mesalamine use in favor of the control group.
Leuven Cohort Risk Factors Table 2 shows the variables that remained statistically significant after stepwise backward elimination in the multivariate Cox regression analysis in the Leuven cohort. Disease extent, PSC, postinflammatory polyps, age at diagnosis, DALM, and flat high-grade dysplasia showed an independent positive association with CRC. The use of anti–TNF-a medication showed a strong negative association.
Prediction Rule After the exclusion of protective factors, medication use, and all types of dysplasia, the following variables remained statistically significant after stepwise elimination and subsequent entering together in a multivariate Cox regression analysis: IBD type, PSC, microscopic
January 2015
Predicting CRC in IBD
151
Table 1. Patient Characteristics and Distributions Between CRC Cases and Controls in the Leuven and Dutch Cohorts Variable
Leuven CRC (n ¼ 50)
Leuven controls (n ¼ 136)
Dutch CRC (n ¼ 138)
Dutch controls (n ¼ 206)
Ulcerative colitis Male sex Age at IBD diagnosis, ya Disease duration, ya Smoking, yesb Family history of CRC Disease extent 50% PSC PIP DALM ALM Flat LGD Flat HGD Colonic stenosis Surveillancec Colectomyd Mesalamine use Thiopurine use Methotrexate use TNF-a antagonist use
60% 60% 28 21 22% 5% 88% 14% 56% 8% 2% 10% 6% 26% 26% 18% 85% 36% 6% 17%
50% 40% 25 17 27% 7% 52% 5% 33% 0% 6% 2% 2% 18% 20% 27% 90% 64% 27% 57%
65% 60% 27 21 19% 9% 85% 9% 71% 21% 24% 28% 13% 31% 27% 22% 77% 29% 3% 8%
51% 53% 26 22 31% 4% 57% 2% 43% 0% N/A N/A N/A 14% 29% 44% 88% 32% 1% 5%
ALM, adenoma-like lesion or mass; HGD, high-grade dysplasia; LGD, low-grade dysplasia; N/A, not available; PIP, postinflammatory polyps. a Age and disease duration are presented as the median. b Smoking behavior was scored as positive for active smoking at the end of the follow-up period or a positive smoking history. c Surveillance was defined as colonoscopy with random biopsy specimens taken every 10 cm of the colon. d Colectomy was defined as any type of surgery in which (a part of) the colon was removed (ranging from ileocecal resection to subtotal colectomy).
disease extent, and the presence of postinflammatory polyps. The results were identical for backward and forward selection elimination processes. In this process, age at diagnosis was no longer statistically significantly associated with CRC, whereas IBD type and disease extent were. Table 3 shows corresponding bs and resultant integer points for the prediction rule. We calculated a total score for each individual patient by adding up the integer points of the characteristics that were present in each patient. This total score per patient had a C-statistic of 0.75 (95% confidence interval, 0.67–0.84). Figure 1 shows the distribution of CRC within the Leuven cohort for each score for Crohn’s disease and ulcerative colitis, respectively. Scores of 0, 13, 23, 27, and 37 represent patients with Crohn’s Table 2. Hazard Ratios for Risk Factors That Are Associated Independently With Developing CRC After Multivariate Cox Regression Analysis in the Leuven Cohort Patient variable
HR
95% CI
Microscopic extent 50% PSC PIP DALM fHGD TNF-a antagonist use Age at IBD diagnosis, >37 y
2.9 3.8 3.0 7.0 6.2 0.2 2.3
0.98–8.4 1.4–10.4 1.6–5.8 1.8–27.0 1.5–26.1 0.1–0.5 1.1–5.0
CI, confidence interval; fHGD, flat high-grade dysplasia; HR, hazard ratio; PIP, postinflammatory polyps.
disease, and scores of 15, 25, 28, 38, 42, and 52 represent patients with ulcerative colitis. Patients with ulcerative colitis were found to have a higher risk of developing CRC. As shown in Figure 1A, Crohn’s disease patients without PSC, without postinflammatory polyps, and with limited disease extent developed no CRC in the Leuven cohort. The highest risk of CRC was found in patients with ulcerative colitis, PSC, postinflammatory polyps, and extensive disease (Figure 1B). Figure 2 shows the same distribution in the Dutch cohort based on the prediction score derived from the Leuven cohort. Again, the lowest proportion (12%) of CRC cases was found in patients with Crohn’s disease with limited disease extent, without PSC or postinflammatory polyps. Patients from the Dutch cohort with ulcerative colitis in whom all risk factors were present had the highest CRC risk, which was as high as 80%. We found a C-statistic of 0.67 for the Dutch cohort. The Leuven and Dutch cohorts both showed an increasing percentage of CRC cases with higher scores. A detailed description of each score with corresponding CRC distributions is shown in Table 4. Based on these data we suggest stratifying IBD patients into low- and high-risk groups of CRC, as shown in Supplementary Figure 1.
Discussion In this study, we show that colitis patients at high or low risk for IBD-associated CRC can be distinguished
152
Lutgens et al
Clinical Gastroenterology and Hepatology Vol. 13, No. 1
Table 3. Prediction Scores in Integer Points Derived From Corresponding Regression Coefficients for Each Patient Characteristic Patient characteristic
b
HR
95% CI
P
b/0.813
Integer
PIP Microscopic disease extent 50% PSC IBD-type ulcerative colitis
0.813 1.070 1.127 1.232
2.3 2.9 3.1 3.4
1.2–4.1 1.0–8.3 1.3–7.3 1.8–6.4
.008 .045 .010 .000
1 1.316 1.386 1.515
10 13 14 15
b, regression coefficient; CI, confidence interval; HR, hazard ratio; PIP, postinflammatory polyps.
using an externally validated prediction rule. These results expand the findings from well-known individual risk factor studies4,5,7,8,10,17 and support expert opinion of current surveillance guidelines to stratify patients in different risk groups. The prediction rule is based on 4 well-established risk factors (ie, IBD-type, PSC, postinflammatory polyps, and disease extent). Our data show that in the presence of all these risk factors the risk of developing CRC increases significantly. Conversely, this means that in the absence of these risk factors (ie,
A Rule for Determining Risk of Colorectal Cancer in Patients With Inflammatory Bowel Disease Maurice Lutgens,*,‡ Séverine Vermeire,§ Martijn Van Oijen,*,‡,k Frank Vleggaar,* Peter Siersema,* Gert van Assche,‡ Paul Rutgeerts,‡ and Bas Oldenburg,* on behalf of the Dutch Initiative on Crohn and Colitis *Department of Gastroenterology and Hepatology, University Medical Center Utrecht, Heidelberglaan, The Netherlands; ‡ Veterans Affairs Center for Outcomes Research and Education, and kDivision of Digestive Diseases, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, California; and §Department of Gastroenterology, University Hospital Gasthuisberg, Leuven, Belgium BACKGROUND & AIMS:
Surveillance guidelines for inflammatory bowel disease–associated colorectal cancer (IBD-CRC) are based on findings from retrospective studies. We aimed to create and validate a prediction rule to assist clinicians in identifying patients with IBD who are at low and high risk for CRC.
METHODS:
We performed a retrospective case-control study of 2 cohorts of patients from tertiary care centers (the University Hospital of Leuven, Belgium, and 7 University Medical Centers in The Netherlands). Multivariate Cox regression was used to select independent risk factors for CRC in the Leuven cohort. Based on their regression coefficients (b), we created a rule to predict risk for CRC. In validation studies, the predictive strength was tested by C-statistic analysis and then validated externally in the Dutch cohort.
RESULTS:
In total, we identified 50 patients with IBD-CRC (cases) and 136 patients with IBD without CRC (controls) in Leuven, and 138 cases and 206 controls in the Dutch cohort. From the Leuven cohort we created the CRC risk prediction rule based on 4 risk factors: IBD-type ulcerative colitis (b [ 1.2), primary sclerosing cholangitis (b [ 1.1), extent of colonic disease ‡50% (b [ 1.1), and postinflammatory polyps (b [ 0.8). The prediction rule consisted of a total score for each individual patient calculated from the presence or absence of these 4 risk factors. For example, a score of 13 represented patients who had extensive Crohn’s disease without PSC or postinflammatory polyps, who had a 15% likelihood of CRC in the Leuven cohort and a 24% likelihood of CRC in the Dutch cohort. Scores of 0, 13, 23, 27, and 37 represented patients with Crohn’s disease, and scores 15, 25, 28, 38, 42, and 52 represented patients with ulcerative colitis. The total score per patient had a C-statistic of 0.75. In the Dutch cohort this score had a C-statistic of 0.67.
CONCLUSIONS:
Ulcerative colitis, primary sclerosing cholangitis, disease extent ‡50%, and postinflammatory polyps were found to determine risk for CRC in patients with IBD. A surveillance guideline that incorporates the relative weights of these risk profiles would identify patients at risk for CRC more accurately than algorithms in current guidelines.
Keywords: Colon Cancer; Inflammation; Screening; UC.
he risk of colorectal cancer (CRC) in inflammatory bowel disease (IBD) colitis patients is known to be increased,1 although not as much as previously reported.2 It is clear that the majority of IBD patients will never develop this complication. The challenge lies in identifying patients at particularly high risk by using established and reliable risk factors. Generally accepted factors are a previous diagnosis of colonic dysplasia,3 disease duration,2,4 disease extent,4 and primary sclerosing cholangitis (PSC).5,6 Endoscopic features also can assist in identifying high-risk patients. Both the presence
T
of postinflammatory polyps7,8 and colonic strictures7 have been shown to be associated with an increased cancer risk in ulcerative colitis. Conversely, a normal endoscopic appearance reduces the risk to the same level as Abbreviations used in this paper: BSG, British Society of Gastroenterology; CRC, colorectal cancer; DALM, dysplasia-associated lesion or mass; IBD, inflammatory bowel disease; PSC, primary sclerosing cholangitis; TNF, tumor necrosis factor. © 2015 by the AGA Institute 1542-3565/$36.00 http://dx.doi.org/10.1016/j.cgh.2014.06.032
January 2015
the background non-IBD population.7 The recently updated British Society of Gastroenterology (BSG) guideline9 also includes family history of CRC as a risk factor based on the study by Askling et al.10 Despite the confirmed association of these predictive and protective factors with regard to CRC development in IBD, their combined value in predicting which patients are at a low or high risk of CRC has never been validated. Therefore, current recommendations on risk groups and intervals in the BSG and American Gastroenterology Association guidelines9,11 are based on expert opinion. The aim of the present study was to discriminate reliably between low and high risk of IBD-associated CRC based on combined risk factors. To achieve this we created an internally and externally validated, easy-to-use prediction rule for IBD-associated CRC.
Methods Study Design We performed a retrospective case-control study to identify predictive and protective factors for IBDassociated CRC. To validate findings in an external cohort we collected data from 2 separate cohorts, 1 from the University Hospital Leuven in Belgium and 1 from university hospitals in The Netherlands.12 The Leuven cohort was used to build the prediction rule and the Dutch cohort served as an external validation.
Patient Selection: Leuven Cohort We used an International Classification of Diseases, 9th revision, coding search for the diagnoses of IBD and CRC at the University Hospital Gasthuisberg, Leuven, which is a tertiary referral center in Flanders, North Belgium. Search results were available for the period from September 1999 to August 2009. The search yielded 99 results. After an initial check, we identified and excluded 3 patients with isolated ileal Crohn’s disease, 21 patients with unconfirmed IBD, 7 patients with synchronous diagnoses of IBD and CRC, 10 patients with low-grade dysplasia but no CRC, 1 patient with high-grade dysplasia but no CRC, and 7 patients with cancers other than colorectal adenocarcinoma. To expand our cohort we cross-referenced the local electronic patient database with all local pathology reports from September 1990 through June 2011. This yielded 5 additional CRC cases between 1990 and 1999. Controls in Leuven were selected from the Leuven IBD biobank by generating a randomly ordered patient list. Controls then were selected consecutively from the top of the list. Crohn’s patients with only ileal involvement were excluded (n ¼ 12). In total, we collected data from 50 cases and 136 unmatched controls. We used unmatched controls to include as many modifiers of CRC risk as possible.
Predicting CRC in IBD
149
Patient Selection: Dutch Cohort The cohort from The Netherlands included all IBDassociated CRC cases from 1990 to 2006 in tertiary referral centers. The cohort of Dutch patients with IBDassociated CRC has been reported previously.12,13 These patients were selected using the nationwide pathology automated archive (known as PALGA).14 Search terms for colitis and carcinoma with multiple synonyms were used to identify patients with IBD and CRC. Results were screened manually for confirmed IBD and CRC. By using this method we identified 149 IBD-associated CRCs in 7 Dutch university medical hospitals. Eleven patients were excluded because of synchronous diagnoses of IBD and CRC. We selected controls by using identical search terms, but this time we excluded neoplasia. A random number generator was used to select controls in a 1:2 ratio to cases. After manual screening, 94 selected controls were excluded because of isolated ileal Crohn’s disease or an unconfirmed diagnosis of IBD. In total, we collected data from 138 cases and 206 unmatched controls.
Data Collection The starting point of data collection and follow-up evaluation was the date of symptom onset that could be attributed to IBD. This was defined by a persistent change of bowel habits and/or bloody diarrhea and/or continuous abdominal pain followed by a diagnosis of IBD. If no clear onset of symptoms was recorded, the date of IBD diagnosis was used. The end of follow-up evaluation for cases was the date of CRC diagnosis. The end of follow-up evaluation for controls could be any of the following: (1) end of study date, which was October 15, 2011, for the Leuven data, and July 1, 2006, for the Dutch data; (2) date of death by any cause; (3) loss to follow-up evaluation defined by the date of the last known visit to the outpatient clinic; or (4) the date of total colectomy. The following variables were collected for both cohorts: sex; IBD type, defined as ulcerative colitis or Crohn’s colitis (indeterminate colitis cases were analyzed with the ulcerative colitis group); smoking, defined as positive for active smoking at the end of the follow-up evaluation or a smoking history; family history of CRC, defined as any first-degree relative having a diagnosis of CRC; limited disease, defined as microscopic disease extent of less than 50%, and extensive disease, defined as microscopic disease extent equal to 50% or greater of the colonic surface; concurrent PSC; postinflammatory polyps; dysplasia-associated lesion or mass (DALM); adenoma-like lesion; flat low-grade dysplasia; flat high-grade dysplasia; colonic stenosis; any mesalamine use; any thiopurine use; any anti–tumor necrosis factor (TNF)-a use; any methotrexate use; colonoscopic surveillance, defined as a dichotomous variable (as yes when random biopsy specimens were taken every 10 cm or chromoendoscopy was performed;
150
Lutgens et al
as no if not); pretumor colonic resection; and age at IBD diagnosis (divided as 36 y). These age cut-off values were based on age distribution within the Leuven cohort.
Statistical Analysis Overall risk factor analysis in the Leuven cohort. Before establishing the prediction score, we first identified risk and protective factors independently associated with CRC. A multivariate Cox regression analysis was used to address different follow-up times between individual patients. This was followed by stepwise conditional elimination using the following categoric variables: sex, IBD type, smoking behavior, microscopic extent more or less than 50% of the colonic surface, PSC, postinflammatory polyps, DALM, adenoma-like mass, flat low-grade dysplasia, flat high-grade dysplasia, colonic stenosis, any mesalamine use, any thiopurine use, any anti–TNF-a use, any methotrexate use, colonoscopic surveillance, pretumor colonic resection, age younger than 19 years, and age older than 37 years. Entering only these 2 age categories and not the middle category of age 19 to 36 years enabled a comparison between the hazard ratio of the younger and older groups with the middle group. The results of this analysis provided an overview from which we could select factors to use in the prediction score.
Prediction Rule The main goal was to create a prediction rule with a minimal number of risk factors. Hence, we only selected independent factors associated with CRC in the Leuven cohort that would be easy to use in clinical practice. We excluded possible protective factors such as colonic resection, colonoscopic surveillance itself, and medication use. We additionally excluded flat low-grade dysplasia, flat high-grade dysplasia and DALM, because these variables are outcomes of surveillance, regardless of their association with CRC. We used multivariate Cox regression analysis with stepwise elimination to identify the remaining factors that had the strongest association with CRC in the Leuven cohort. Only factors with a final P value of less than .05 were included in the prediction rule. The selected risk factors were assigned weights that were derived from their regression coefficients (b). After division by the smallest b, all bs were converted to integer scores, for example, a b of 2.2 corresponded to 2 points and a b of 14.3 corresponded to 14 points. The prediction rule consisted of the total additive score per patient based on the presence or absence of these risk factors. A higher total score per patient indicated a higher risk of developing CRC. To test its discriminative power, we calculated a C-statistic for the total score of each individual patient.15,16 To this end, we used a Survival Analysis SAS Macro that is publicly available from the
Clinical Gastroenterology and Hepatology Vol. 13, No. 1
Mayo Clinic (http://www.mayo.edu/research/departm ents-divisions/department-health-sciences-research/divi sion-biomedical-statistics-informatics/software/locallywritten-sas-macros). We report C-statistic values based on calculations with ties. Calibration is the degree of correspondence between the estimated probability produced by the score and the actual observed outcomes. We assessed the calibration performance by comparing the predicted risk for CRC with the observed number of CRCs in the Leuven cohort (internal validation) and the Dutch cohort (external validation). SPSS version 20 (Armonk, NY) and SAS version 9 (Cary, NC) software were used for all analyses.
Results Patient Characteristics Table 1 shows the overall distribution of variables in the Leuven and the Dutch cohorts. In both cohorts there were more ulcerative colitis and male patients with CRC. Disease duration was longer in the CRC groups. Active smoking or a smoking history was more prevalent in the control groups. A family history of CRC was slightly more prevalent in the Leuven control group and the Dutch CRC group. Disease extent, PSC, and postinflammatory polyp ratios were similar in the Leuven and Dutch cohorts with more extensive disease, PSC, and postinflammatory polyps than in the CRC groups. DALMs were seen only in the CRC group. The Dutch control group had no dysplasia because those patients specifically were excluded when the data for this cohort were collected. More CRC cases than controls had a colonic stenosis, although a colectomy was more common in controls. Especially in the Leuven cohort, more controls were on thiopurines and methotrexate medication and were treated with anti–TNF-a. Only the Dutch cohort showed a difference in mesalamine use in favor of the control group.
Leuven Cohort Risk Factors Table 2 shows the variables that remained statistically significant after stepwise backward elimination in the multivariate Cox regression analysis in the Leuven cohort. Disease extent, PSC, postinflammatory polyps, age at diagnosis, DALM, and flat high-grade dysplasia showed an independent positive association with CRC. The use of anti–TNF-a medication showed a strong negative association.
Prediction Rule After the exclusion of protective factors, medication use, and all types of dysplasia, the following variables remained statistically significant after stepwise elimination and subsequent entering together in a multivariate Cox regression analysis: IBD type, PSC, microscopic
January 2015
Predicting CRC in IBD
151
Table 1. Patient Characteristics and Distributions Between CRC Cases and Controls in the Leuven and Dutch Cohorts Variable
Leuven CRC (n ¼ 50)
Leuven controls (n ¼ 136)
Dutch CRC (n ¼ 138)
Dutch controls (n ¼ 206)
Ulcerative colitis Male sex Age at IBD diagnosis, ya Disease duration, ya Smoking, yesb Family history of CRC Disease extent 50% PSC PIP DALM ALM Flat LGD Flat HGD Colonic stenosis Surveillancec Colectomyd Mesalamine use Thiopurine use Methotrexate use TNF-a antagonist use
60% 60% 28 21 22% 5% 88% 14% 56% 8% 2% 10% 6% 26% 26% 18% 85% 36% 6% 17%
50% 40% 25 17 27% 7% 52% 5% 33% 0% 6% 2% 2% 18% 20% 27% 90% 64% 27% 57%
65% 60% 27 21 19% 9% 85% 9% 71% 21% 24% 28% 13% 31% 27% 22% 77% 29% 3% 8%
51% 53% 26 22 31% 4% 57% 2% 43% 0% N/A N/A N/A 14% 29% 44% 88% 32% 1% 5%
ALM, adenoma-like lesion or mass; HGD, high-grade dysplasia; LGD, low-grade dysplasia; N/A, not available; PIP, postinflammatory polyps. a Age and disease duration are presented as the median. b Smoking behavior was scored as positive for active smoking at the end of the follow-up period or a positive smoking history. c Surveillance was defined as colonoscopy with random biopsy specimens taken every 10 cm of the colon. d Colectomy was defined as any type of surgery in which (a part of) the colon was removed (ranging from ileocecal resection to subtotal colectomy).
disease extent, and the presence of postinflammatory polyps. The results were identical for backward and forward selection elimination processes. In this process, age at diagnosis was no longer statistically significantly associated with CRC, whereas IBD type and disease extent were. Table 3 shows corresponding bs and resultant integer points for the prediction rule. We calculated a total score for each individual patient by adding up the integer points of the characteristics that were present in each patient. This total score per patient had a C-statistic of 0.75 (95% confidence interval, 0.67–0.84). Figure 1 shows the distribution of CRC within the Leuven cohort for each score for Crohn’s disease and ulcerative colitis, respectively. Scores of 0, 13, 23, 27, and 37 represent patients with Crohn’s Table 2. Hazard Ratios for Risk Factors That Are Associated Independently With Developing CRC After Multivariate Cox Regression Analysis in the Leuven Cohort Patient variable
HR
95% CI
Microscopic extent 50% PSC PIP DALM fHGD TNF-a antagonist use Age at IBD diagnosis, >37 y
2.9 3.8 3.0 7.0 6.2 0.2 2.3
0.98–8.4 1.4–10.4 1.6–5.8 1.8–27.0 1.5–26.1 0.1–0.5 1.1–5.0
CI, confidence interval; fHGD, flat high-grade dysplasia; HR, hazard ratio; PIP, postinflammatory polyps.
disease, and scores of 15, 25, 28, 38, 42, and 52 represent patients with ulcerative colitis. Patients with ulcerative colitis were found to have a higher risk of developing CRC. As shown in Figure 1A, Crohn’s disease patients without PSC, without postinflammatory polyps, and with limited disease extent developed no CRC in the Leuven cohort. The highest risk of CRC was found in patients with ulcerative colitis, PSC, postinflammatory polyps, and extensive disease (Figure 1B). Figure 2 shows the same distribution in the Dutch cohort based on the prediction score derived from the Leuven cohort. Again, the lowest proportion (12%) of CRC cases was found in patients with Crohn’s disease with limited disease extent, without PSC or postinflammatory polyps. Patients from the Dutch cohort with ulcerative colitis in whom all risk factors were present had the highest CRC risk, which was as high as 80%. We found a C-statistic of 0.67 for the Dutch cohort. The Leuven and Dutch cohorts both showed an increasing percentage of CRC cases with higher scores. A detailed description of each score with corresponding CRC distributions is shown in Table 4. Based on these data we suggest stratifying IBD patients into low- and high-risk groups of CRC, as shown in Supplementary Figure 1.
Discussion In this study, we show that colitis patients at high or low risk for IBD-associated CRC can be distinguished
152
Lutgens et al
Clinical Gastroenterology and Hepatology Vol. 13, No. 1
Table 3. Prediction Scores in Integer Points Derived From Corresponding Regression Coefficients for Each Patient Characteristic Patient characteristic
b
HR
95% CI
P
b/0.813
Integer
PIP Microscopic disease extent 50% PSC IBD-type ulcerative colitis
0.813 1.070 1.127 1.232
2.3 2.9 3.1 3.4
1.2–4.1 1.0–8.3 1.3–7.3 1.8–6.4
.008 .045 .010 .000
1 1.316 1.386 1.515
10 13 14 15
b, regression coefficient; CI, confidence interval; HR, hazard ratio; PIP, postinflammatory polyps.
using an externally validated prediction rule. These results expand the findings from well-known individual risk factor studies4,5,7,8,10,17 and support expert opinion of current surveillance guidelines to stratify patients in different risk groups. The prediction rule is based on 4 well-established risk factors (ie, IBD-type, PSC, postinflammatory polyps, and disease extent). Our data show that in the presence of all these risk factors the risk of developing CRC increases significantly. Conversely, this means that in the absence of these risk factors (ie,
