If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Optimizing Screening for Colorectal Cancer: An Algorithm Combining Fecal Immunochemical Test, Blood-Based Cancer-Associated Proteins and Demographics to Reduce Colonoscopy Burden
Gastro Unit, Hvidovre Hospital, Hvidovre, DenmarkInstitute for Clinical Medicine, University of Copenhagen, Copenhagen, DenmarkDepartment of Surgery, Nordsjællands Hospital, Hillerød, Denmark
Department of Public Health Programmes and University Research Clinic for Cancer Screening, Randers Regional Hospital, Randers, DenmarkDepartment of Clinical Medicine, Aarhus University, Aarhus, Denmark
Colorectal cancer screening using Fecal Immunochemical Test (FIT) reduces both morbidity and mortality.
•
However, FIT screening is linked to high demand of colonoscopy capacity mainly due to a high false positive rate.
•
In here, we tested whether our proposed algorithm using FIT, blood-based biomarker analysis and demographics of the subject could improve screening.
•
The algorithm reduced the amount of needed colonoscopies by 4%-11%.
Abstract
Background
Fecal Immunochemical Test (FIT) is widely used in population-based screening for colorectal cancer (CRC). This had led to major challenges regarding colonoscopy capacity. Methods to maintain high sensitivity without compromising the colonoscopy capacity are needed. This study investigates an algorithm that combines FIT result, blood-based biomarkers associated with CRC, and individual demographics, to triage subjects sent for colonoscopy among a FIT positive (FIT+) screening population and thereby reduce the colonoscopy burden.
Materials and methods
From the Danish National Colorectal Cancer Screening Program, 4048 FIT+ (≥100 ng/mL Hemoglobin) subjects were included and analyzed for a panel of 9 cancer-associated biomarkers using the ARCHITECT i2000. Two algorithms were developed: 1) a predefined algorithm based on clinically available biomarkers: FIT, age, CEA, hsCRP and Ferritin; and 2) an exploratory algorithm adding additional biomarkers: TIMP-1, Pepsinogen-2, HE4, CyFra21-1, Galectin-3, B2M and sex to the predefined algorithm. The diagnostic performances for discriminating subjects with or without CRC in the 2 models were benchmarked against the FIT alone using logistic regression modeling.
Results
The discrimination of CRC showed an area under the curve (AUC) of 73.7 (70.5-76.9) for the predefined model, 75.3 (72.1-78.4) for the exploratory model, and 68.9 (65.5-72.2) for FIT alone. Both models performed significantly better (P < .001) than the FIT model. The models were benchmarked vs. FIT at cutoffs of 100, 200, 300, 400, and 500 ng/mL Hemoglobin using corresponding numbers of true positives and false positives. All performance metrics were improved at all cutoffs.
Conclusion
A screening algorithm including a combination of FIT result, blood-based biomarkers and demographics outperforms FIT in discriminating subjects with or without CRC in a screening population with FIT results above 100 ng/mL Hemoglobin.
Colorectal Cancer (CRC) is globally the third leading cause of cancer worldwide. The total number of new cases in 2018 was approximately 1.850.000 (10.2% of all cancers) and it was responsible for 881.000 deaths.
CRC is a preventable disease if early detection and removal of a pre-cancerous (adenoma) or low CRC stage is successful. As methods to reduce the burden of CRC is desirable, population-based screening modalities have already been implemented. Most population-based screening methods are stool-based tests such as the Fecal Immunochemical Test (FIT). Those screened with a stool-based test have a reported 25% reduction in CRC mortality.
Beside reducing CRC incident cases and mortality a population-based screening program with FIT might also be cost-effective mainly because long and expensive cancer treatments are averted.
When implementing a screening program using a stool-based approach such as FIT, cut-off levels of stool-based tests must be considered. The quantitative cutoff level distinguishing between a positive and negative test result, and ultimately whether a person should subsequently be offered a follow up colonoscopy, determines both the sensitivity and specificity. Lowering the cutoff levels leads to higher sensitivity (ie identification of more true positives) while lowering the specificity (ie increase of more false positives) and vice versa. In other words, high detection of CRC and adenomas using FIT is inextricably linked to high demand of colonoscopy capacity. One study suggests that the optimal cutoff in terms of balanced sensitivity and specificity is 100 ng/mL Hemoglobin
(leading to 89% sensitivity and 91% specificity). However, a recent study suggest that the optimal cutoff value in a FIT population-based screening program is 45 ng/mL Hemoglobin when sensitivity and specificity are weighted equally.
Even small percentage changes in specificity will have a substantial impact on the needed colonoscopy capacity, since population-based screening settings often include hundreds of thousands of people.
Though evidence and recommendations suggest FIT cutoff values between 45 and 100 ng/mL Hemoglobin to reduce CRC incidence and at the same time be cost-effective, many countries are not adhering.
Recommendations on Fecal Immunochemical Testing to Screen for Colorectal Neoplasia: a consensus statement by the US Multi-Society Task Force on Colorectal Cancer.
. The difference between using a cutoff of 100 ng/mL compared to 500 ng/mL results in missing approximately 40% of the cancers using the higher cutoff.
The use of higher cut-offs is likely caused by the lack of resources allocated for colonoscopies. Therefore, methods to maintain high sensitivity without compromising the colonoscopy capacity are warranted. This might be accomplished by combining blood-based biomarker tests and FIT results. A systematic review
regarding this combination found that blood tests might enhance sensitivity but only by decreasing specificity. However, the majority of the studies were not conducted on true screening cohorts and overestimation and underestimation of sensitivities and specificities, respectively, compared to a screening cohort is plausible.
This study aimed to develop and test if an algorithm combining the FIT result, blood-based biomarkers associated with CRC, and demographics, as previously suggested,
can outperform the discrimination of FIT alone (utilizing a FIT cutoff of ≥100 ng/mL hemoglobin) in a true screening population. If so, such an algorithm could be a better and viable alternative than simply using a higher FIT cutoff to select individuals for follow-up colonoscopies.
Methods
This study was based on a 2-step screening approach; it included only individuals with a positive FIT result (≥100 ng/mL Hemoglobin), who were then subsequently offered a blood test. A predefined algorithm and an exploratory algorithm (see below) were benchmarked against the performance of FIT only (FIT model).
The study has been reported according to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement.
Protocol outlines for parts 1 and 2 of the prospective endoscopy III study for the early detection of colorectal Cancer: validation of a concept based on blood biomarkers.
(approved by the Scientific and Ethical Committee at the Capital Region of Denmark (H-4-2013-050) and the GDPR-aligned Danish Data Protection Agency (2007-58-0015/HVH-2013-022). In brief, the Endoscopy III study included both screening positive (FIT+) and negative (FIT−) individuals from the Danish Screening Program for Colorectal Cancer. The screening program is population-based with biennial invitations for all individuals aged 50 to 74 and participation is completely free of charge. It uses a FIT (OC sensor, Eiken Pharmaceuticals, Japan) with a detection range of ≤35 to ≥1000 ng/mL Hemoglobin. FIT result of ≥100 ng/mL will lead to a subsequent offer of colonoscopy.
Endoscopy III study was initiated March 2014 at 9 collaborating hospitals in 2 different regions and ended late August 2016 (30 months). Participants were randomly selected from the screening program and offered to participate in the study. FIT+ subjects were only included in the study if they accepted subsequent colonoscopy since knowledge of colonoscopy-findings and subsequent pathology results on resected tissue were required. Just prior to the colonoscopy, FIT+ participants were verbally asked to participate in the study, sign an informed consent and donate blood samples. The Endoscopy III study part 1 was divided into a “Training Set” and a “Validation Set” based on chronological time. The Validation Set was formed by counting backwards until 200 CRCs were obtained among the FIT+ participants. The Training Set included 4,048 FIT+ subjects, including 242 CRCs, who were eligible for this study (Figure 1).
Figure 1Flow chart of the study population from the Endoscopy III part 1 project. FIT+ individuals from the Training set were included (N = 4,048).
Cancer stage are accordingly to the Union for International Cancer Control stage (UICC-stage).
Demographics, medical history, findings and interventions at colonoscopy or subsequent operation, and the pathology of removed tissue were entered into a Web-based database. The database was audited to verify that all data records were correct. Analyses of blood samples were blinded from outcomes and results of blood sample analyses were subsequently merged with the clinical data.
Predictors
Demographics (age and sex), the FIT result and 9 blood-based biomarkers (Carcinoembryonic antigen [CEA], high-sensitivity C-reactive protein [hsCRP], Tissue inhibitor of metalloproteinase-1 [TIMP-1], Pepsinogen-2, Human epididymis protein 4 [HE4], Cytokeratin 19 fragment [CyFra21-1], Galectin-3, Ferritin and Beta-2-microglobulin [B2M]) were evaluated as potential predictors.
Outcomes
The primary outcome was the discrimination of CRC vs. non-CRC individuals, while the secondary outcome was the discrimination of CRC/HRA/MRA vs. LRA or clean colorectum.
Presence of CRC or adenomas were defined by the histopathology report of removed lesions by either the colonoscopy or resected tissue by following surgery. All resected adenomas (only regarding tubulo-, tubulovillous-, and villous adenomas, traditional serrated adenomas, or sessile serrated adenomas with dysplasia) were risk-stratified for each patient and allocated into 3 categories, according to the European guideline for quality assurance in colorectal cancer screening and diagnosis (First Edition)
: high-risk adenomas (HRA), medium-risk adenomas (MRA), and low-risk adenomas (LRA). High-risk allocation included: one lesion ≥20mm, or presence of ≥5 adenomas regardless of size, or removal by piecemeal approach. Medium-risk allocation included: one lesion ≥10mm - <20mm, or 3-4 adenomas regardless of size, or tubulovillous/villous adenoma, or presence of high-grade neoplasia. Low-risk allocation included: 1 to 2 lesions <10mm, or tubular adenomas, or low-grade neoplasia. Clean colorectum was defined as the absence of malignancy or adenomas at colonoscopy.
Blood Sample Collection, Processing, Handling, and Storage
All blood samples were collected just prior to colonoscopy according to a validated Standard Operating Procedure, by trained employees to minimize of preanalytical variation. Blood draw was performed from an antecubital vein into 10 mL K2 EDTA plasma tubes (Becton-Dickinson Vacutainer, NJ). The tubes were centrifuged at 3,000 xg for 10 min at 21°C, leaving plasma, buffy coats and erythrocytes separated. Plasma were transferred to a 50 mL Nunc centrifugation tube and re-centrifuged at 3,000 g for additional 10 min at 21°C. Subsequently, plasma was transferred to MAXXLINE cryo tubes with preprinted unique barcodes (Dacos A/S, Esbjerg, Denmark) and frozen at -80°C. All samples were marked with unique barcodes for easy storage identification, which was performed using FreezerWorks PC-based tracking system (FreezerWorks, Seattle, WA). The entire procedure was completed within 2 hours.
Biomarker Analyses
Cancer-related biomarkers were selected based on previous results and literature review.
Rationale for the 9 biomarkers selected can be viewed in Supplemental Table 1. CEA, TIMP-1, Pepsinogen-2, HE4, CyFra21-1, Galectin-3, and Ferritin levels were investigated using the chemiluminescence microparticle immunoassay approach and the ARCHITECT i2000 automated platform according to manufacturer's instructions (Abbott Laboratories, Abbott Park, IL) at the Abbott Center of Excellence at Amsterdam UMC.
In brief, 20-100 µl of sample were mixed with 5µM paramagnetic microparticles coated with antibodies against CEA, TIMP-1, Pepsinogen, HE4, CyFra21-1, Galectin 3, and Ferritin. Following an incubation time of 18 minutes the samples were washed and acridinium-labelled monoclonal antibody conjugate binding the same proteins added in the previous step. Following an incubation of 4 minutes and a second wash, reaction mixture containing hydrogen peroxide at alkaline pH was added to the tubes and the chemiluminescent signal from the acridinium was recorded from each captured antigen. hsCRP and B2M protein levels were measured using immunoturbidimetric technology and Clinical Chemistry assays (ARCHITECT c8000, Abbott Laboratories, Abbott Park, IL) in which anti-CRP or anti-B2M antibody-coated latex particles were mixed and agglutinated to the CRP or B2M proteins with the EDTA samples. The agglutination was detected as an absorbance change (at 572 nm) with the rate of absorbance change being proportional to the quantity of protein.
Statistics
The sample size of the current study was predefined by the sample size calculation for the Endoscopy III study,
Protocol outlines for parts 1 and 2 of the prospective endoscopy III study for the early detection of colorectal Cancer: validation of a concept based on blood biomarkers.
which was based on previously published prescreening studies and estimated that inclusion of 8000 FIT+ would result in approximately 450 (5,6%) – 500 (6,3%) identified with CRC.
Data were imported into R statistical software version 4.1.0
from which all statistical analyses were performed. All available data was used, and no imputation of missing data was performed (3 patients had missing data for B2M, of whom 1 had additional missing data for Galectin-3 and TIMP-1). Continuous data were presented as median and quartiles. Categorical data were presented as numbers and frequencies (%). Log base 10 transformation was applied for all biomarker values. The primary outcome was the discrimination of CRC vs. non-CRC individuals, while the secondary outcome was the discrimination of CRC/HRA/MRA vs. LRA or clean colorectum. Logistic regression models were used to provide predictions of the dichotomous outcomes.
Two different algorithms were tested: A predefined model including FIT result, age, and clinically available biomarkers and an exploratory model using less established biomarkers. For the predefined model, 3 biomarkers (CEA, Ferritin and hsCRP) were chosen for their well-known biological involvement in CRC, their low cost regarding analytic expenses, and their implementation feasibility as these markers are ready-to-use in a clinical setting.
For the exploratory model we added potential predictors, such as sex, TIMP-1, Pepsinogen-2, HE4, CyFra21-1, Galectin-3, and B2M, in conjunction with FIT result, age, CEA, hsCRP, and Ferritin. For the exploratory model, the strongest predictors were chosen using backwards regression retaining exploratory variables with a P-value ≤.05, which resulted in different models for the primary and secondary outcomes. Both models were benchmarked against the FIT model only including the FIT result. The FIT model was modified to include cutoff levels of ≥100, ≥200, ≥300, ≥400, and ≥500 ng/mL Hemoglobin.
To assess the performance of the 3 prediction models the area under the receiver operating characteristic (ROC) curve (AUC) and the Brier score
were calculated for each model. Calibration was inspected graphically. Validation of the models were done using 10-fold cross-validation repeated 10 times. Brier score was reported after multiplication by 100.
Sensitivity, specificity, positive predictive value, negative predictive value, true positives, false positives, false negatives, true negatives, number needed to colonoscopy, and number of colonoscopies per event were calculated for all models and the additive effect of the predefined and explanatory models was investigated in relation to the FIT models using the same number of true positives and false positives for FIT cutoffs at ≥100, ≥200, ≥300, ≥400, and ≥500 ng/mL. P-values ≤.05 were considered statistically significant.
Results
Among the 4,048 FIT+ participants from the Danish Colorectal Cancer Screening Program, 242 (6%) subjects were diagnosed with CRC and 2,064 (51%) with adenomas categorized into 548 (27%) HRA, 812 (39%) MRA and 704 (34%) LRA (Figure 1). The remaining 1,742 (43%) subjects presented with a clean colorectum, which included other non-malignant conditions such as diverticula, hemorrhoids, hyperplastic polyps and other inflammatory polyps (Figure 1). The study population comprised of 2,433 (60.1%) men with slight differences between the subgroups (CRC, adenoma and clean colorectal). The overall median age was 65 years, also with a small difference between patients with CRC or adenoma and healthy individuals. In total, 459 (11,3%) individuals reported they had had a previous cancer (excluding basal cell skin cancer). The number of individuals having CRC with a FIT result between 100 to 200 ng/mL Hemoglobin was 2.3% and 13.5% for subjects with highest measurable FIT of ≥1000 ng/mL (Table 1).
Table 1Study Population Characteristics by Subgroups
Total (n = 4048)
Clean Colorectum (n = 1742)
Adenomas (n = 2064)
Colorectal Cancer (n = 242)
Sex
Male
2433 (61%)
870 (50%)
1408 (68%)
155 (64%)
Female
1615 (39%)
872 (50%)
656 (32%)
87 (36%)
Age, y
Median(Q1, Q3)
65.2 (57.4-71.0)
62.6 (54.5-69.9)
66.4 (59.3-71.8)
67.7 (62.6-73.5)
Previous cancer
459 (11%)
187 (11%)
238 (12%)
34 (14%)
FIT value, ng/mL
100-199
1329 (33%)
702 (40%)
596 (29%)
31 (13%)
200-299
695 (17%)
321 (18%)
344 (17%)
30 (12%)
300-399
395 (10%)
169 (10%)
208 (10%)
18 (7%)
400-499
238 (6%)
95 (6%)
134 (6%)
9 (4%)
500-999
520 (13%)
198 (11%)
285 (14%)
37 (15%)
≥1000
871 (21%)
257 (15%)
497 (24%)
117 (48%)
Continuous data presented as median and quartiles (Q1 = 1st quartile 25%, Q3 3rd quartile 75%). Categorical data presented as numbers and frequencies (%).
For the primary outcome, both the predefined- and exploratory model performed significantly better (P < .001) than the FIT model in the discrimination of cases with CRC vs. cases with no CRC with an AUC of 73.7 (70.5-76.9) and 75.3 (72.1-78.4) vs. 68.9 (65.5;72.2) (Figure 2). The exploratory model included the predictors age, sex, FIT, CEA, Ferritin, hsCRP, HE4 and Cyfra21-1, and performed significantly better than the predefined model (with age, FIT, CEA, Ferritin, and hsCRP) with an increase in the AUC by 1.6 (0.2-3.0) (P < .001).
Figure 2ROC curves for the FIT, predefined, and exploratory models for the primary outcome (CRC vs. non-CRC).
In order to ensure that the predefined and exploratory models were reliable, a 10-fold cross-validation repeated 10 times were conducted. This resulted in slightly lower AUCs for the predefined, exploratory and FIT models of 72.7 (64.9-80.4), 74.0 (65.9-82.1) and 67.3 (60.2-74.3) (Table 2).
Table 2Overview of the Multivariable Analysis for the Primary and Secondary Outcome
The calibration plots (Figure 3) comparing the predicted probability of CRC vs. the actual fraction of persons having CRC showed a maximum estimated risk of CRC using the FIT model was only 13.4%, which corresponds to a FIT result of ≥1000 ng/mL.
Figure 3Calibration plot (plots of estimated probabilities against the observed binary outcome) and Brier scores for the FIT, predefined, and exploratory models for the primary outcome. The closer the plots are to the diagonal line the better the model. The lower Brier score the better.
For the 2 algorithms, the maximum predicted probability of CRC was higher and visually both the predefined and exploratory algorithms appeared more accurate in giving individual risk predictions (plots closer to the diagonal line, particularly from 0 to 25% risk prediction). Also, the 2 algorithms showed significantly (P < .001) lower Brier scores (predefined model 5.3 [4.7-5.9] and exploratory model 5.2 [4.6-5.7]) compared to the FIT model Brier score (5.5 [4.9-6.1]).
Identification of CRC, HRA or MRA Cases
For the secondary outcome, discrimination of patients with CRC, HRA or MRA vs. individuals with LRA or no CRC, both the predefined and exploratory (now including age, sex, FIT, CEA and B2M) models performed better than the FIT model with an AUC of 66.7 (65.0-68.4) and 68.0 (66.3-69.7) vs. the FIT AUC of 65.0 (63.2-66.7) (P < .001) (Figure 4). After cross-validation the AUCs for the predefined, exploratory and FIT models decreased to 66.1 (64.2-67.9), 67.4 (65.5-69.3) and 64.1 (62.4-65.8) (Table 2).
Figure 4ROC curves for the FIT, predefined, and exploratory models for the secondary outcome (CRC, HRA, MRA vs. LRA or clean colorectum).
Impact on Required Colonoscopies Using Different Screening Modalities
To investigate the impact on colonoscopies needed in a screening modality adding blood-based biomarkers to the FIT screening in a 2-step approach (Figure 5) instead of simply increasing the FIT cutoff, we benchmarked the predefined and exploratory models against FIT cutoffs at 100, 200, 300, 400, and 500 ng/mL Hemoglobin. We used fixed sensitivities and specificities corresponding to the same numbers of true positives and false positives in all analyses. For both the primary- and secondary outcome, addition of blood-based biomarkers and demographics to the FIT screening increased the number of true negatives in all analyses at fixed true positives and decreased the number of false negatives in all analyses at fixed false positives (Table 3 and 4).
Figure 5Proposed two-step screening algorithm.
Most standard screening method rely solely on the FIT result resulting in a high demand of follow up colonoscopies. We propose an algorithm-based, two-step screening approach, with an initial FIT test followed up by a blood test if the FIT is positive. Based on the combined result of the blood test, FIT result and demographics (age and sex), an algorithm decides whether colonoscopy is required. We hypothesized that this screening method could lower the number of unnecessary colonoscopies in healthy individuals while still correctly allocating individuals with colorectal neoplasia to further clinical follow-up.
Table 3Performance of the Predefined- and Exploratory Models Benchmarked Against the FIT Model at Chosen cutoffs (≥100, ≥200, ≥300, ≥400 and ≥500 ng/mL) for Primary Outcome (CRC vs. non-CRC) and at Fixed True Positives (TP) and False Positives (FP) (Bolded and Italicized Numbers)
Outcome And Models
Sensitivity
Specificity
PPV
NPV
TP
FP
FN
TN
Needed Colono-scopies
Colono-scopy per cancer
CRC vs. non-CRC
FIT model ≥100ng/mL
100%
0%
6.0%
100%
242
3806
0
0
4048
16.7
Pre. model
100%
5%
6.3%
100%
242
3603
0
203
3845
15.9
Explor. model
100%
3%
6.2%
100%
242
3688
0
118
3930
16.2
FIT model ≥200ng/mL
87.2%
34.1%
7.8%
97.7%
211
2508
31
1298
2719
12.9
Pre. model
87.2%
36.7%
8.1%
97.8%
211
2408
31
1398
2619
12.4
Pre. model
90.0%
34.1%
8.0%
98.2%
218
2508
24
1298
2726
12.5
Explor. model
87.2%
41.8%
8.7%
98.1%
211
2215
31
1591
2426
11.5
Explor. model
91.7%
34.1%
8.1%
98.5%
222
2508
20
1298
2730
12.3
FIT model ≥300ng/mL
74.8%
51.6%
8.9%
96.9%
181
1843
61
1963
2024
11.2
Pre. model
74.8%
59.6%
10.6%
97.4%
181
1537
61
2269
1718
9.5
Pre. model
80.6%
51.6%
9.6%
97.7%
195
1843
47
1963
2038
10.5
Explor. model
74.8%
58.9%
10.4%
97.4%
181
1565
61
2241
1746
9.6
Explor. model
78.9%
51.6%
9.4%
97.5%
191
1843
51
1963
2034
10.6
FIT model ≥400ng/mL
67.4%
61.5%
10.0%
96.7%
163
1466
79
2340
1629
10.0
Pre. model
67.4%
67.7%
11.7%
97.0%
163
1231
79
2575
1394
8.6
Pre. model
71.9%
61.5%
10.6%
97.2%
174
1466
68
2340
1640
9.4
Explor. model
67.4%
69.5%
12.3%
97.1%
163
1162
79
2644
1325
8.1
Explor. model
73.6%
61.5%
10.8%
97.3%
178
1466
64
2340
1644
9.2
FIT model ≥500ng/mL
63.6%
67.5%
11.1%
96.7%
154
1237
88
2569
1391
9.0
Pre. model
63.6%
70.0%
11.9%
96.8%
154
1142
88
2664
1296
8.4
Pre. model
67.4%
67.5
11.7%
97.0%
163
1237
79
2569
1400
8.6
Explor. model
63.6%
72.8%
13.0%
96.9%
154
1034
88
2772
1188
7.7
Explor. model
69.4%
67.5%
12.0%
97.2%
168
1237
74
2569
1405
8.4
Colonoscopy per cancer refers to the number of colonoscopies needed to be performed to find one cancer.
Table 4Performance of the Predefined and Exploratory Models Benchmarked Against the FIT Model at Chosen Cutoffs (≥100, ≥200, ≥300, ≥400 and ≥500 ng/mL Hemoglobin) for the Secondary Outcome (CRC, HRA or MRA vs. LRA or Clean Colorectum) and at Fixed True Positive (TP) Value and False Positive (FP) Value (Bolded and Italicized Numbers)
Outcome And Models
Sensitivity
Specificity
PPV
NPV
TP
FP
FN
TN
Needed Colono-scopies
Colonoscopy per CRC, HRA, MRA
CRC, HRA or MRA vs. LRA or clean
FIT model ≥100ng/mL
100%
0%
36.9%
100%
1494
2554
0
0
4048
2.7
Pre. model
100%
0.4%
37.0%
100%
1494
2539
0
15
4033
2.7
Explor. model
100%
1.5%
37.3%
100%
1494
2511
0
40
4005
2.7
FIT model ≥200ng/mL
78.7%
39.6%
43.3%
76.1%
1176
1543
318
1011
2719
2.3
Pre. model
78.7%
41.5%
44.0%
75.9%
1176
1494
318
1060
2670
2.3
Pre. model
79.6%
39.6%
43.5%
76.8%
1189
1543
305
1011
2732
2.3
Explor. model
78.7%
42.3%
44.1%
77.2%
1176
1472
318
1079
2648
2.3
Explor. model
81.5%
39.5%
44.1%
78.4%
1217
1543
277
1008
2760
2.3
FIT model ≥300ng/mL
63.8%
58.0%
47.1%
73.3%
953
1071
541
1483
2024
2.1
Pre. model
63.8%
59.4%
47.9%
73.7%
953
1038
541
1516
1991
2.0
Pre. model
64.5%
58.1%
47.4%
73.7%
964
1071
530
1483
2035
2.1
Explor. model
63.8%
63.1%
50.3%
74.8%
953
941
541
1610
1894
1.9
Explor. model
68.4%
58.0%
48.8%
75.8%
1022
1071
472
1480
2093
2.0
FIT Model ≥400ng/mL
54.4%
68.0%
49.8%
71.8%
812
817
682
1737
1629
2.0
Pre. model
54.4%
69.2%
50.8%
72.2%
812
786
682
1768
1598
1.9
Pre. model
55.5%
68.0%
50.4%
72.3%
829
817
665
1737
1646
1.9
Explor. model
54.4%
71.5%
52.8%
72.8%
812
727
682
1824
1539
1.9
Explor. model
58.7%
68.0%
51.8%
73.8%
877
817
617
1734
1694
1.9
Fit model ≥500ng/ml
48.4%
73.8%
52.0%
71.0%
723
668
771
1886
1391
1.9
Pre. model
48.4%
75.1%
53.2%
71.3%
723
635
771
1919
1358
1.8
Pre. model
49.4%
73.8%
52.5%
71.4%
738
668
756
1886
1406
1.9
Explor. model
48.4%
76.1%
54.2%
71.6%
723
610
771
1941
1333
1.8
Explor. model
51.8%
73.8%
53.7%
72.3%
774
668
720
1883
1442
1.8
Colonoscopy per cancer refers to the number of colonoscopies needed to be performed to find one cancer.
Focusing, for example, on the primary outcome with a FIT cutoff 200 ng/mL, the FIT model missed 31 out of 242 CRCs (Table 3). Compared to the predefined model and the exploratory model at fixed false positive numbers, the number of missed CRCs were 24 and 20, respectively – thereby salvaging 23% and 35% of FIT-missed CRCs. At the fixed true positive number, the number of needed colonoscopies for the predefined model and the exploratory model was 2,619 and 2,426, respectively, vs. 2,719 in the FIT model, to find the same number of CRC. The reduction in needed colonoscopies was therefore 3.7% and 10.8%, respectively (Table 3). This pattern was evident throughout all the different FIT screening modalities.
Discussion
With this study, including 4,048 FIT+ individuals, we developed 2 screening algorithms (a predefined model and an exploratory model) including a combination of age, sex, 6 blood-based biomarkers (CEA, hsCRP, Ferritin, HE4, CyFra21-1 and B2M), and the FIT results, which improved the prediction of CRC and CRC/HRA/MRA compared to a screening modality using the FIT result alone. Both the predefined- and the exploratory model for the discrimination of subjects with CRC vs. subjects without CRC significantly outperformed the FIT screening modality with AUCs of 73.7 (70.5-76.9) and 75.3 (72.1-78.4), respectively, compared to 68.9 (65.5-72.2) for the FIT screening alone.
The FIT-based screening modality has been implemented in several western countries, but as approximately 40% of the screening colonoscopies are without any neoplastic findings, new screening strategies are required in order to reduce the number of required colonoscopies. A systematic review by Niedermaier et al.
in 2018 described 8 studies combining a FIT result with blood-based biomarkers for early colorectal neoplasia detection. They concluded a limited gain by combination of blood tests with the FIT, which are likely dependent on the choice of biomarkers selected. A study by Symonds et al.
in 1,381 individuals, combining FIT (at cutoff 300 ng/mL hemoglobin) in combination with 2 DNA methylation markers (BCAT1 and IKZF1), showed a 82% sensitivity with a specificity of 87% concerning discrimination of CRC vs. non-CRC individuals. In comparison, our study found a sensitivity of 81% and specificity of 52% for the predefined model and 79% sensitivity and 52% specificity for the exploratory model, respectively (Table 3). Another study by Werner et al.
investigating 1,660 individuals used 5 blood biomarkers (CEA, Ferritin, Seprase, Osteopontin and anti-p53 antibody) in combination with FIT (cutoff not reported) and reported a sensitivity of 83% (95% CI 61-95) at 92% (95% CI 90-94) specificity. The increased sensitivity and specificity in the above-mentioned studies may be induced by cohort design, which were quite different and likely influence the assessed performance. The first study was a mixture of both symptomatic and asymptomatic patients and the latter a screening cohort (invitation to a free screening colonoscopy) and not a FIT+ screening cohort, making it difficult to compare their results with the current study. In a third study including 216 retrospectively selected patients undergoing colonoscopy, 5 fecal inflammatory biomarkers were investigated.
In this study, 1 biomarker, M2-pyruvate-kinase, was found to improve the sensitivity of the FIT screening (≥50 ng/mL Hemoglobin) from 91.5% to 96.6%. However, the specificity was compromised from 72.3% to 58.2% and could not be increased without decreasing sensitivity. In contrast, a study with similar cohort design utilizing 341 FIT+ individuals investigated the CRC predictive value of the fecal biomarkers calprotectin, lactoferrin, transferrin alone or in addition to fecal Hemoglobin and did not find any improvement in the diagnostic accuracy in the combined test.
The variation in the diagnostic accuracy may likely be affected by the biology of the biomarkers selected with tumor-derived DNA fragments in the blood being more pronounced in late-stage tumors, while inflammatory biomarkers may be more suitable for early detection of colorectal neoplasia as the immune system may be more sensitive to these early alterations. Though the inflammatory biomarkers investigated in this study are associated with several other inflammatory diseases, the combination of such biomarkers with the more CRC-specific fecal Hemoglobin levels holds great potential to improve the diagnostic accuracy.
The calibration plots that compare the predicted probability of CRC vs. the actual fraction of persons having CRC differed between the predefined- and exploratory models compared to FIT model (Figure 3). Two features are notably different: 1) The maximum estimated risk of CRC using FIT model was only 13,4%, which corresponds to a FIT result of ≥1000 ng/mL. For the 2 algorithms, the maximum predicted risk was higher, however after 25% to 30% the risk prediction curves flatline and should not be interpreted further 2) Both the predefined- and exploratory algorithms visually appear more accurate in providing individual risk predictions (plots are closer to the diagonal line), which is also reflected by the lower Brier scores (a Brier score equal to 0 means perfect prediction and a brier score of 100 is equal to a model that is completely inaccurate). An improved and higher individual risk prediction could aid subjects in deciding whether or not to accept subsequent colonoscopy or guide clinical staff regarding proper follow-up.
In Denmark the threshold for FIT positivity is set at ≥100 ng/mL. However, due to limited colonoscopy capacities, other countries have higher cut off resulting in a decreased sensitivity for colorectal neoplasia. A systematic review and meta-analysis by Selby et al.
have investigated over 64 studies regarding the effect of changing FIT cutoffs and describe similar and resembling results as our study (Tables 3 and 4). Both algorithms in this study were benchmarked against FIT alone using the number of true positives and false positives for relevant FIT cutoffs used in population-based screening (ranging from ≥100 to ≥500 ng/mL Hemoglobin) and all performance metrics were improved in both models at all cutoffs. These results may encourage countries with constraints in the colonoscopy capacity to switch approaches and consider using this 2-step screening approach so that either less cancers are missed without increased demands in colonoscopy capacity or decrease the colonoscopy capacity without increasing missed cancers.
Considering the secondary outcome, which was the discrimination of CRC, HRA or MRA vs. individuals with LRA or clean colorectum, the predictor variables for the exploratory model were quite different than for the primary endpoint. The only beneficial biomarkers included appeared to be CEA and B2M. The different biomarker selections for the primary and secondary outcome was not unexpected due to the molecular differences in the adenoma-carcinoma sequence and the known phenomena that the ability of biomarkers to detect adenomas is somewhat weak.
It could therefore be expected to be increased in those individuals with premalignant lesions such as HRA or MRA, as a sign of inflammation. However, we saw the inverse effect in this study. The odds ratio was 0.645 suggesting that B2M is downregulated for individuals with either CRC, HRA or MRA compared to those with LRA or clean colorectum. These results were strengthened by the study by Bednarz-Mias et al.
, who likewise found a lower expression of B2M for patients with CRC compared to controls and increased levels of B2M in patients with inflammatory bowel disease. B2M is a subunit of the Human Leucocyte Antigen (HLA) type I receptor on all cells and is required for antigen presentation of host cell DNA in order to avoid being killed by the immune cells in the host. However, while normal cells expressing normal DNA will be spared, this immune surveillance mechanism will inevitably kill tumor cells expressing mutated DNA fragments. Thus, to avoid the immune system, which may be present at increased levels during inflammatory bowel disease, the tumors need to downregulate their B2M/HLA expression removing the presentation of their mutated DNA sequences. This may be a very early stage evasion mechanism, which is more pronounced in adenomas than cancers, as loss of B2M may limit the tumors ability to metastasize.
Beta-2-microglobulin mutations are linked to a distinct metastatic pattern and a favorable outcome in microsatellite-unstable stage IV gastrointestinal Cancers.
combining FIT (cutoff 300 ng/mL) and 2 methylation markers (BCAT1 and IKZF1) reported a sensitivity of 27% at 87% specificity for detection of advanced adenoma (which corresponds to HRA and MRA adenomas in our study). Another study by Tao et al.
which, similar to our study, used logistic regression models for FIT in combination with 3 inflammation biomarkers (CRP, sCD26 and TIMP-1) showed an AUC of 73% in detection of advanced adenomas compared to 68% using the FIT model alone. Like our study, combining the blood-based biomarkers with FIT did not lead to any truly meaningful enhancement in detection of adenomas and further research in biomarkers specific for these early neoplastic lesions are still warranted.
The annual report of The Danish Colorectal Cancer Screening Database contains detailed information of all FIT+ participants and the results of the colonoscopy. In the same timespan as our study, 6.2% (2014), 5.8% (2015) and 5.9% (2016) of participants of the Danish Screening Program for Colorectal Cancer had CRC.
Gastric Cancer screening by combined determination of serum helicobacter pylori antibody and pepsinogen concentrations: ABC method for gastric Cancer screening.
A strength of our study is the close resemblance in the numbers of cancers in the present cohort vs. all screened individuals in Denmark in that period, suggesting that this study population is representative of the target screening population. The 3 biomarkers CEA, hsCRP and Ferritin included in the predefined model were chosen because of their association to colorectal malignancy, but also for their known easily available clinical use and low cost. If the algorithm is to be applied as the standard screening approach (see Figure 5) in population-based screening in countries using FIT, the consequences of adding a blood test in the screening must be investigated in terms of impact on compliance. It may have a negative effect on the compliance by lowering the clinical sensitivity of the screening, since people might be discouraged by spending further time and transport to have a blood test conducted, and thus having the opposite effect of that intended. On the other hand, if the method has a higher test performance and more accurately allocates an individual for subsequent colonoscopy, it might encourage individuals overall to participate in the screening program as they may avoid an unnecessary and unpleasant examination.
In conclusion, the results of the study indicate that usage of a screening algorithm including a combination of FIT result, blood-based biomarkers and demographics may increase detection of CRC without the need to increase the colonoscopy capacity or reduce the total amount of needed colonoscopies without compromising CRC detection compared to FIT screening. The results achieved in this study must be considered an early phase toward development of an optimized screening algorithm. Further development using additional blood-based biomarkers and further validation of the models must be achieved before implementing the suggested algorithm screening setting.
Clinical Practice Points
•
Colorectal cancer (CRC) screening using Fecal Immunochemical Test (FIT) reduces both morbidity and mortality, but high detection of CRC and precancerous adenomas using FIT is inextricably linked to high demand of colonoscopy capacity.
•
Improved selection to colonoscopy and thus a reduction in the false positive rate are advantageous for the colonoscopy capacity, the healthcare budget and the subjects undergoing an unnecessary, potential unpleasant and risk-associated procedure.
•
We proposed an algorithm-based, 2-step screening approach: Initial FIT test was followed up by a blood test if the FIT was positive (cutoff value of 100 ng/mL Hemoglobin). Three entities were used in the algorithm: FIT, blood-based biomarkers and demographics (age and sex) and decided whether colonoscopy was necessary.
•
The main results for the algorithm compared to sole use of FIT discovered an overall reduction in colonoscopy requirements by 4%-11% without reducing detection of colorectal cancer.
•
An overall improvement of CRC screening by usage of the algorithm could become clinically relevant in situations where colonoscopy capacity constraints are present or for countries that are using or are considering using FIT cutoffs over 100 ng/mL Hemoglobin.
Disclosure
Gerard Davis, Susan Gawel and Frans Martens are all employees of Abbott Laboratories Inc. All other authors declare that they have no competing interest.
Acknowledgment
We thank the patients for their participation and the research nurses, secretaries and technicians at the participating hospitals for their skillful work with subject recruitment, blood collection and data recording. This work was supported by The Andersen Foundation, The Augustinus Foundation, The Beckett Foundation, The Inger Bonnéns Fund, The Hans & Nora Buchards Fund, The Walter Christensen Fund, The P.M. Christiansen Fund, The Kong Chr. X's Fund, The Aase & Ejnar Danielsens Fund, The Family Erichsens Fund, The Knud & Edith Eriksens Fund, The Svend Espersens Fund, The Elna & Jørgen Fagerholts Cancer Research Foundation, The Sofus Carl Emil Friis Scholarship, The Torben & Alice Frimodts Fund, The Eva & Henry Frænkels Fund, The Gangsted Foundation, Thora & Viggo Groves Memorial Scholarship, The H-Foundation, Erna Hamiltons Scholarship, Søren & Helene Hempels Scholarship, The Sven & Ina Hansens Fund, The Henrik Henriksen Fund, Carl & Ellen Hertz’ Scholarship, Jørgen Holm & Elisa F. Hansen Memorial Scholarship, The Jochum Foundation, The KID Foundation, The Kornerup Foundation, The Linex Foundation, The Dagmar Marshalls Fund, The Midtjyske Fund, The Axel Muusfeldts Fund, The Børge Nielsens Fund, Michael Hermann Nielsens Memorial Scholarship, The Arvid Nilssons Fund, The Obelske Family Fund, The Krista & Viggo Petersens Fund, The Willy & Ingeborg Reinhards Fund, The Kathrine & Vigo Skovgaards Fund, The Toyota Foundation, The Vissing Foundation, The P. Carl Petersen Fund, The Danish Cancer Research Foundation, Hvidovre Hospitals Research Pool.
Recommendations on Fecal Immunochemical Testing to Screen for Colorectal Neoplasia: a consensus statement by the US Multi-Society Task Force on Colorectal Cancer.
Protocol outlines for parts 1 and 2 of the prospective endoscopy III study for the early detection of colorectal Cancer: validation of a concept based on blood biomarkers.
Beta-2-microglobulin mutations are linked to a distinct metastatic pattern and a favorable outcome in microsatellite-unstable stage IV gastrointestinal Cancers.
Gastric Cancer screening by combined determination of serum helicobacter pylori antibody and pepsinogen concentrations: ABC method for gastric Cancer screening.
00006-3&id=gr1.jpg){kind=link}
00006-3&id=gr2.jpg){kind=link}
00006-3&id=gr3.jpg){kind=link}
00006-3&id=gr4.jpg){kind=link}
00006-3&id=gr5.jpg){kind=link}