Abstract
Background/Aim: This study assessed the diagnostic accuracy (DA) of the predictive features of colorectal cancer (CRC, predictCRC), triage process (triage), and ColonView (CV) fecal immunochemical test (FIT) in a CRC screening setting. The diagnostic score models (DMs) including predictCRC with triage and CV test were also calculated. Patients and Methods: The study cohort of 544 patients included 58 CRC patients and 486 non-CRC patients who submitted three consecutive fecal samples for analysis, by two fecal occult blood (FOB) assays (CV FIT test, HemoccultSENSA test). Hierarchical multilevel logistic models were used to test the DA (for CRC) of each item of predictCRC (with triage I and II) and DMs, visualized as hierarchical summary receiving operating characteristic (HSROC) curves. Results: The DA of the predictCRC location of neoplasm (Loc), triage I, and triage II showed 49%, 41%, and 93% sensitivity (Se), and 70%, 99.5%, and 88% specificity (Sp), respectively. The PPV+ of triage I (92%) was higher than that of Loc (22%) or triage II test (45%). In the conventional receiver operating characteristic (ROC) analysis, the area under the curve (AUC) values for the different DMs ranged from 0.880 (for DM without triage I and II), whereas the highest AUC value of 0.960 was reached for DM with triage I and II included in the formula. In the HSROC analysis, the AUC values were as follows: i) with all predictCRCs, AUC=0.717 and ii) with DMs, AUC=0.937. In the roccomp analysis, the difference in AUC values between i) and ii) was statistically significant (p<0.0001). Conclusion: In the detection of CRC, the DA of the new DMs with triage was far superior to that of DMs without triage. This is the first study to report evidence of improved DA in the detection of CRC using DMs including predictCRC with triage and CV FIT test.
Abdominal pain (Abd), intestinal/rectal bleeding (Bleed), a change in bowel habits (Hab), anemia, and weight loss are widely considered to be alarming predictive features for colorectal cancer (predictCRC), although these are also common symptoms for other intestinal diseases. However, most patients with these symptoms do not have any of these diseases. The fecal immunochemical tests (FITs) for hemoglobin are commonly used as triage tests, but little is known about the predictCRCs and other alarming symptoms in CRC screening settings.
In the literature, a few reports assessing the diagnostic accuracy (DA) of FITs to triage patients with predictCRCs have been published (1, 2). Some evidence of a high DA is shown in the reviews by Westwood et al. (3), D’Souza et al. (4), and Booth et al. (5), who assessed the DA of predictCRCs and showed that FITis are superior to symptoms alone (3). Unfortunately, previous investigations, meta-analyses, and trials (1-5) did not include ColonView FIT (CV test). Meklin et al. (6, 7) investigated the DA of the CV test and reported that the DA of the automatic reading mode (AR) was higher than that of the visual reading mode (VR) in CRC detection. In further analyses (8, 9), the authors showed that both VR and AR modes reached higher DAs for distal CRCs than for proximal CRCs. In addition, these authors assessed the DA of the CV test among the bleed-negative CRCb− and bleed-positive CRCb+ patients demonstrating that the AR mode showed significantly better DA as compared to the VR reading in CRCb+ and CRCb− groups (10, 11).
One of the obvious weaknesses of the predictCRC assessment is that during consultations or in a screening setting, all the symptoms experienced by patients are rarely recorded. The combination of FITs and alarming symptoms may be potentially helpful in the triage of patients in CRC screening, because the detection of occult blood by FITs usually precedes alarming symptoms by 2-3 years. This study attempted to investigate the DA of the predictCRC assessment, triage as a part of the diagnostic process, as well as CV FIT test in detecting CRC in a screening setting. No previous studies have estimated the DA of these diagnostic approaches using hierarchical multilevel logistic (HSROC) models.
Patients and Methods
Study design. The Barretos Cancer Hospital (BCH) colorectal neoplasia (CRN) cohort included 5,090 patients between January 2014 and December 2016. Detailed descriptions of all CRN lesions were recorded, including their number and size as well as predictCRC during the colonoscopy consultation. The final study cohort of 544 patients included 58 patients with CRC and 486 patients without CRC (12).
Methods. Two optional CV techniques are available: visual reading (VR) mode and automatic reading (AR) mode (12). For the AR mode, the Quick Test Reader (QTR) is needed, and the protocol of this instrument has recently been described in more detail (Chembio Diagnostics GmbH, Berlin, Germany) (12). The guaiac-based FOBT (Hemoccult SENSA, Beckman Coulter Inc., Pasadena, CA, USA) was used for comparison.
First triage (triage I) was performed at colonoscopy consultation, including assumption of ‘possible signs of intestinal bleeding’ and second triage (triage II) was performed after colonoscopy including assumption of ‘colonoscopy finding to match with fecal occult blood test (FOBT) result’. Triages I and II included a suggestion of possible diagnosis for patients with CRC.
DM models. In a multivariate logistic (stepwise) regression analysis, SPSS software was used (SPSS statistics 26.0.0.1; IBM, Armonk, NY, USA) and the clinical features shown in Table I and Table II were included in the stepwise analysis as binary data, for example, CRC=1 and non-CRC=0. Using the coefficients of the stepwise model, a DM was built and its predictive value for CRC was estimated.
Different study variables for colorectal cancer endpoint.
Diagnostic score models (DMs) for colorectal cancer (CRC) screening. The DMs shown at five different combinations of the variables shown in Table I. Cut-off levels: DM I=−2.09, DM II=−2.13, DM III=−2.07, DM IV=−1.52, DM V=−1.43.
1. DM for CRC without the SENSA test and without triage I/II. (PE=positive endpoint and NE=negative endpoint).
DM=0.61 × Sex (PE=Male=1, NE=Female=0) + 0.78 × Age (PE=1, NE=0) – 0.70 × Weight (PE=1, NE=0) + 0.93 × Abdominal pain (PE=1, NE=0) – 2.29 × Hemorrhoids (PE=1, NE=0) + 1.73 × CV Hb AR (PE=1, NE=0) + 2.31 × CV Hb/Hp AR (PE=1, NE=0) + 1.01 × Location (PE=1, NE=0) – 4.28 (=constant).
2. DM for CRC without Triage I/II.
DM=−1.02 × Age (PE=1, N=0) + 0.79 × Abdominal pain (PE=1, NE=0) – 2.28 × Hemorrhoids (PE=1, NE=0) + 0.99 × SENSA (PE=1, NE=0) + 1.61 × CV Hb AR (PE=1, NE=0) + 1.83 × CV Hb/Hp AR (PE=1, NE=0) + 1.05 × Location (PE=1, NE=0) – 4.67 (=constant).
3. DM for CRC without Triage II.
DM=−1.20 × Weight (PE=1, N=0) – 5.30 × Triage I (PE=1, NE=0) + 1.04 × Location (PE=1, NE=0) + 1.82 × CV Hb AR (PE=1, NE=0) + 2.26 × CV Hb/Hp AR (PE=1, NE=0) – 3.64 (=constant).
4. DM for CRC without Triage I.
DM=−0.63 × Weight (PE=1, NE=0) + 0.75 × Abdominal pain (PE=1, NE=0) + 1.22 × Location (PE=1, NE=0) – 2.28 × Hemorrhoids (PE=1, NE=0) + 2.44 × CV Hb/Hp AR (PE=1, NE=0) + 4.10 × Triage II (PE=1, NE=0) – 5.68 (=constant).
5. DM for CRC, when all features were included.
DM=−1.13 × Weight (PE=1, NE=0) + 1.00 × Location (PE=1, NE=0) + 3.55 × Triage I (PE=1, NE=0) + 3.40 × Triage II (PE=1, NE=0) +1.60 × CV Hb AR (PE=1, NE=0) + 1.74 × CV Hb/Hp AR (PE=1, NE=0) – 5.19 (=constant).
Statistical methods. STATA/SE version 17.0 (StataCorp, College Station, TX, USA) and MetaDiSc software 1.4 (Unit of Clinical Biostatistics team of the Ramón y Cajal Hospital, Madrid, Spain) were used for analysis of the data as previously reported (6-11, 13). Estimating the clinical usefulness of predictCRCs, FOBTs, and DMs, we have reported the usefulness index (UI) (14-16) for different variables shown in Table I and five different combinations for DMs shown in Table II. Conventional receiver operating characteristic (ROC) analysis was used to determine the optimal cut-off (CO) values for both Hb and Hb/Hp of the CV test. Meta-analytical technique (metaprop; Stata) was used to create separate forest plots for sensitivity (Se) and specificity (Sp), with each set of data included. The hierarchical multilevel logistic regression models were used to create hierarchical summary ROC (HSROC) curves using the CRC endpoint. Roccomp test (Stata) was used to compare the area under the curve (AUC) values of the HSROC curves.
Results
DA of sex, age, weight, and body mass index (BMI). The Se, Sp, and efficiency (Ef) of the sex as a test for CRC were 52%, 62%, and 60%, respectively (Table I). Conventional ROC analysis was used to determine the optimal cut-off (CO) values for age, weight, and BMI as test for CRC. The Se, Sp, and efficiency (Ef) of the age as a test for CRC were 55%, 71%, and 69%, respectively (Table I, Figure 1, and Figure 2). The Se, Sp, and Ef of the weight and BMI as tests for CRC were 51%/33%, 32%/53%, and 34%/51%, respectively (Table I, Figure 1, and Figure 2). The positive predictive value (PPV+) of age as a test was higher than that of the weight or BMI: 18% versus 8%/8%, respectively (Table I). The UI values of sex, age, weight, and BMI ranged between −0.09 and 0.14 (Table I).
Sensitivity values of different predictive variables (predictCRCs) for colorectal cancer endpoint. ES: Estimated sensitivity; CI: confidence interval.
Specificity values of different predictive variables (predictCRCs) for colorectal cancer endpoint. ES: Estimated specificity; CI: confidence interval.
DA of the alarming CRC symptoms; abdominal pain (Abd), change in bowel habits (Hab), and rectal bleeding (Rect). The most sensitive symptoms: Abd and Hab showed 59% Se (Table II, Figure 1), while the most specific symptom in CRC diagnosis, Rect, showed a Sp of 69% (Figure 2). The PV+ of the Rect as test was slightly higher than that of the Abd or Hab: 16% versus 14%/14%, respectively (Table I). The Abd, Hab, and Rect showed considerably low UI values of 0.09, 0.09, and 0.08, respectively (Table I).
DA of the main disease status; diverticular disease (Div), hemorrhoids (Hem), angiodysplasia (Ang), and inflammatory bowel disease (Ibd). The Se, Sp, and Ef of the Div as CRC test were 44%, 64%, and 60%, respectively (Table II, Figure 1, and Figure 2). The Se, Sp, and Ef of the Hem, Ang, and Ibd for CRC were 2%/0%/0%, 90%/96%/99%, and 80%/86%/88%, respectively (Table II, Figure 1, and Figure 2). The PV+ of the Div was substantially higher than that of Hem, Ang or Ibd test (Table II): 9% versus 2%/0%/0%. The UI values of Div, Hem, Ang, and Ibd were very low ranging between 0 and 0.02 (Table I).
DA of the location of neoplasm (Loc), triage I and triage II. The DA of the location of neoplasm (Loc), Triage I, and Triage II showed 49%/41%/93% Se. The overall Sp of Loc, triage I, and triage II for the CRC endpoint were 70%, 99.5%, and 88%, respectively. The PPV+ of the Triage I was clearly higher than that of Loc or Triage II test: 92% versus 22%/45%. The UI value of triage II (UI=0.73) was higher than that of triage I (UI=0.41) or Loc (UI=0.30) (Table I).
DA of the CV test VR and AR modes. The most sensitive CV VR test modes: Hb VR and Hb/Hp VR showed 93%/93% Se in the diagnosis of CRC. The Sp, Ef, and PV+ of the Hb VR and Hb/Hp VR were 59%/59%, 62%/62%, and 20%/19%, respectively. In CV AR mode, the conventional ROC analysis showed the optimal CO value of >117 for CV Hb AR (Table II) and >78 for CV Hb/Hp AR (Table II). Using these COs, the Se, Sp and Ef of the CV Hb AR (Table II) and CV Hb/Hp AR (Table II) tests for CRC were 80%/83%, 92%/90%, and 91%/89%, respectively. The PV+ was slightly higher in CV Hb AR test than that in CV Hb/Hp AR test (Table II); 54% versus 48%.
Conventional ROC and AUC values. The Se values for different DMs ranged from 0.74 (for DM I) to 0.93 (for DM IV) (Figure 3), whereas the Sp values for different DMs ranged from 0.86 (for DM II) to 0.94 (for DM V) (Figure 4). ROC analysis was used to determine the optimal CO values for the AR CV test. In addition, conventional ROC curves were used to report the pooled overall DA and AUC values of predictCRC and DMs in CRC. In the conventional ROC analysis, the AUC values for different DMs ranged from 88% to 96%. The DMs without Triage I and II (DM II) showed the lowest AUC values, whereas the highest AUC value of 96% for DM V was reached with all the predictCRCs with triage I and II included in the formula.
Sensitivity values of different diagnostic models (DMs) for colorectal cancer endpoint. ES: Estimated sensitivity; CI: confidence interval.
Specificity values of different diagnostic models (DMs) for colorectal cancer endpoint. ES: Estimated specificity; CI: confidence interval.
HSROC and AUC values. HSROC curves were used to report the pooled overall DA of predictCRC and DM modes. In the HSROC analysis, the AUC values for i) predictCRCs, and ii) DMs were as follows: i) AUC=0.717 (95%CI=0.652-0.782) (Figure 5) ii) and AUC=0.937 (95%CI=0.910-0.964) (Figure 6). In roccomp analysis, the difference in AUC values between i) and ii) was statistically significant (p<0.0001).
Hierarchical summary receiver operating characteristic (HSROC) curve of the different predictive variables (predictCRCs) for colorectal cancer endpoint.
Hierarchical summary receiver operating characteristic (HSROC) curve of the different diagnostic models (DMs) for colorectal cancer endpoint.
Discussion
The National Bowel Cancer (NBC) report from the UK (17) showed that, more than 55% of all CRC patients were diagnosed following a general practitioner (GP) referral, 20% following an emergency admission (EAd) and only 9% were diagnosed through the CRC screening (data missing in 16%). CRC treatment with curative intent was possible more often in patients diagnosed through screening (90%) or following GP referral (70%) than those, who presented with EAd (52%) (17). Therefore, we should boost the CRC screening and awareness of CRC symptoms to reduce the proportion of EAds. It has been suggested that using FITs to select patients for referral has the potential to reduce unnecessary colonoscopies and provide more accurate classification of these patients than does the traditional, symptoms-based guidelines (17-19).
The FIT for hemoglobin (Hb) detects occult blood in the feces, which was not visible to the naked eye, using antibodies specific to the globin moiety of human Hb, either qualitatively or quantitatively. The principles of the FIT assay for human Hb were described by Suovaniemi, who investigated an antibody specific to human globin, the protein component of Hb (20). The most recent FIT test (ColonView®, CV) has the advantage of detecting two components of fecal occult blood (FOB): Hb and hemoglobin/haptoglobin (Hb/Hp) complex (12).
This study reported for the first time the DA of i) the predictCRCs, ii) triage process at endoscopy, iii) CV FIT test with Hb and Hp, as well as five DM models were compared among 58 CRC and 486 non-CRC patients, included in a cohort of 5,090 patients. In the detection of CRC, the DA (AUC-value) of the DMs with triage I + CV Hb AR + CV Hb/Hp AR (DM III, AUC=0.95) or with triage II + CV Hb/Hp AR (DM IV, AUC=0.95) or both triage I and II included with CV Hb AR + CV Hb/Hp AR (DM V, AUC=0.96) proved to be higher to DMs without triage (DM I and DM II). The DM (DM V) with highest AUC value included body weight (PE=1, NE=0), Loc (PE=1, NE=0) + triage I (PE=1, NE=0) + triage II (PE=1, NE=0) +CV Hb AR (PE=1, NE=0) + CV Hb/Hp AR (PE=1, NE=0). These results indicate that DA in detection of CRC can be improved using DMs including predictCRC with triage and CV FIT test.
We have identified two large systematic reviews estimating the accuracy of alarming CRC symptoms and FITs (3, 21). The review by Jellema et al. (21) included three studies on quantitative FITs used to examine symptomatic and asymptomatic patients. Sp was consistently high for family history (0.91; 95%CI=0.75-0.98), weight loss (0.89; 95%CI=0.72-0.96), and iron deficiency anemia (0.92; 95%CI=0.83-0.95); however, all tests suffered from low Se. The authors concluded that although combinations of symptoms and FIT results showed acceptable DA for CRC, evidence from the clinical use of this combination is insufficient (21).
Westwood et al. (3) included 10 studies evaluating the results of FITs as predictors of CRC in colonoscopy referral patients. This review (3) included four FIT assays for Hb and one FIT assay for human Hb/Hp-complex, which are currently available in Europe. These data indicate that a negative FIT result reliably ruled out CRC in approximately three-quarters of the subjects with low-risk bowel symptoms, thus offering an opportunity to reduce the number of unnecessary colonoscopies. The authors concluded that FITs are an effective strategy for investigating the patients with alarming symptoms because they are at high risk for CRC (3). Unfortunately, this analysis did not include the ColonView FIT test.
Meklin et al. (22, 23) reviewed FOBTs used in CRC screening and demonstrated that FITs have a significantly higher DA as compared with traditional guaiac-based FOBTs (gFOBTs). In a formal meta-regression included in their meta-analysis, the authors did not show firm evidence of a superiority of any test brand. Of all qualitative FITs, ColonView-FIT, InstantView and Prevent ID seemed to be the top three choices for CRC screening. Meklin et al. (6, 7) investigated the DA of the CV test reporting that the DA of the AR mode is higher than the VR mode in detecting CRC. In further analyses (8, 9), the authors showed that both VR and AR modes reached higher DAs for distal CRCs than for proximal CRCs. The same authors also demonstrated that the AR mode showed significantly better DA as compared to the VR reading in both CRCb+ and CRCb− (10, 11).
In the present study, in conventional ROC analysis, the AUC values for different DMs ranged from 0.88 (for DM without triage I and II), whereas the highest AUC value of 0.960 was obtained for DM with triage I and II included in the formula. In the HSROC analysis, the AUC values were as follows: i) with all predictCRCs, AUC=0.717 and ii) with DMs, AUC=0.937. The difference in AUC values between i) and ii) was statistically significant (p<0.0001). Thus, the results of ROC and HSROC analyses support a strong link between triage process, CV FIT modes and DMs in CRC diagnosis. In detection of CRC, the DA of the new DMs with triage proved to be far superior to DMs without triage. As pointed out before, this is the first study to report an improved DA in the detection of CRC, by using DMs that include predictCRC with triage and CV FIT test.
We have identified only two studies assessing the DA of the prediction models (DM) that include alarming CRC symptoms and FIT results. Rodríguez-Alonso et al. (24) conducted a multivariate analysis, using logistic regression attempting to assess independent predictors of CRC and advanced neoplasia. The CRC analysis identified male sex (OR=2.39, 95%CI=1.039-5.519; p=0.041), iron-deficiency anemia (OR=2.99, 95%CI-1.27-7.03; p=0.012) and faecal Hb ≥10 μg Hb/g feces (OR=86.60, 95%CI=11.70=641.16; p<0.001) as independent predictors. None of the clinical two-week referral criteria was identified as an independent predictor in the DM, that included fecal Hb measured by FIT. Unfortunately, the authors did not compare the AUC values available, and the study did not include CV FIT test (24).
As part of the COLONPREDICT study, Cubiella et al. (25) tried to develop a prediction model (called the FAST score) for CRC in symptomatic patients, based on the fecal Hb, age and sex. In their score, age was included as a continuous variable and fecal Hb as a categorical variable (0, 0-20, 20-200, and ≥200 μg Hb/g feces). The validation cohort for this model used data from four studies (24, 26-28) and an additional cohort was recruited to the COLONPREDICT study. A weakness in the latter cohort is that the FIT was measured using many different test brands. In the validation cohort, a FAST score of ≥4.5 had a sensitivity of 89.3% (95%CI=84.1-93.0%) and a specificity of 82.3% (95%CI=81.1-83.5%) for CRC.
Although, alarming CRC symptoms have been suggested to predict CRC in some studies (29, 30), our models including certain risk factors of CRC: sex, age, weight, BMI, Abd, Hab, Rect, IBD, angiodysplasia, colonic diverticulum or hemorrhoids; did not provide evidence that CRC diagnosis can be confirmed solely on the basis of alarming CRC features or existing risk factors. However, Loc and triage at endoscopy improved the DA of CRC detection. Some authors have attempted to develop DMs (24, 25) to refine the DA of CRC diagnosis, but these DMs have not yet been fully validated and do not include the CV FIT test or triage process at endoscopy.
No previous reports have investigated the usefulness index (UI) of predictCRCs, FOBTs and DMs to triage the patients with alarming CRC symptoms. Estimating the clinical usefulness of predictCRCs, FOBTs and DMs, we have calculated the UI for different variables shown in Table I and five different combinations for DMs shown in Table II. UI runs coherently from −1 to 1 and tests where the UI is over 0.35 are regarded as useful (14-16). In the UI analysis of this study, 0.48/0.47 UI values were reached for the VR mode of the CV Hb/CV Hb/Hp test, and the UI values for the AR mode of the CV Hb/CV Hb/Hp test were even higher (0.57/0.60). The UI values for the five DMs ranged between 0.46 and 0.78, DM IV and DM V reaching the highest UI values 0.78/0.69, respectively. These data indicate that the UI analysis supports a major impact of both CV FIT (AR, VR) modes and DMs in diagnosis of CRC.
Conclusion
This is the first investigation to report the DA of the AR and VR techniques of the CV test with alarming CRC symptoms in the DMs. This analysis is important for confirming the strength of the AR mode of CV FIT with alarming symptoms in CRC detection. Our study is also the first to assess the DA of the simple prediction score as a triage test for CRC screening patients with symptoms and emphasizes the utility of including FIT results as part of a simple risk score.
Acknowledgements
The study was funded by the North Savo Regional Fund (Pohjois-Savon Maakuntarahasto).
Footnotes
Authors’ Contributions
All Authors contributed to the collection and analysis of data, drafting, and revising the manuscript, and read and approved the final article.
Conflicts of Interest
The Authors have no conflicts of interest or financial ties to disclose in relation to this study.
- Received September 23, 2023.
- Revision received October 18, 2023.
- Accepted October 19, 2023.
- Copyright © 2023 International Institute of Anticancer Research (Dr. George J. Delinasios), All rights reserved.
This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY-NC-ND) 4.0 international license (https://creativecommons.org/licenses/by-nc-nd/4.0).
